bcachefs: Break up io.c

More reorganization, this splits up io.c into
 - io_read.c
 - io_misc.c - fallocate, fpunch, truncate
 - io_write.c

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

34 files changed, 1751 insertions, 1692 deletions

diff --git a/fs/bcachefs/Makefile b/fs/bcachefs/Makefile
index 0a4d2fed66c1..9c00dabb26ac 100644
--- a/fs/bcachefs/Makefile
+++ b/fs/bcachefs/Makefile
@@ -46,7 +46,9 @@ bcachefs-y		:=	\
 	fs-io-pagecache.o	\
 	fsck.o			\
 	inode.o			\
-	io.o			\
+	io_read.o		\
+	io_misc.o		\
+	io_write.o		\
 	journal.o		\
 	journal_io.o		\
 	journal_reclaim.o	\
diff --git a/fs/bcachefs/alloc_foreground.c b/fs/bcachefs/alloc_foreground.c
index e02749ddc362..8e1888a89011 100644
--- a/fs/bcachefs/alloc_foreground.c
+++ b/fs/bcachefs/alloc_foreground.c
@@ -25,7 +25,7 @@
 #include "disk_groups.h"
 #include "ec.h"
 #include "error.h"
-#include "io.h"
+#include "io_write.h"
 #include "journal.h"
 #include "movinggc.h"
 #include "nocow_locking.h"
diff --git a/fs/bcachefs/btree_io.c b/fs/bcachefs/btree_io.c
index 0edbb73a5ec8..00f53cb5d44b 100644
--- a/fs/bcachefs/btree_io.c
+++ b/fs/bcachefs/btree_io.c
@@ -14,7 +14,7 @@
 #include "debug.h"
 #include "error.h"
 #include "extents.h"
-#include "io.h"
+#include "io_write.h"
 #include "journal_reclaim.h"
 #include "journal_seq_blacklist.h"
 #include "recovery.h"
diff --git a/fs/bcachefs/btree_io.h b/fs/bcachefs/btree_io.h
index cd99bbb00a5a..7e03dd76fb38 100644
--- a/fs/bcachefs/btree_io.h
+++ b/fs/bcachefs/btree_io.h
@@ -7,7 +7,7 @@
 #include "btree_locking.h"
 #include "checksum.h"
 #include "extents.h"
-#include "io_types.h"
+#include "io_write_types.h"
 
 struct bch_fs;
 struct btree_write;
diff --git a/fs/bcachefs/compress.c b/fs/bcachefs/compress.c
index 6b17f7cc5860..f1651807c2b7 100644
--- a/fs/bcachefs/compress.c
+++ b/fs/bcachefs/compress.c
@@ -3,7 +3,6 @@
 #include "checksum.h"
 #include "compress.h"
 #include "extents.h"
-#include "io.h"
 #include "super-io.h"
 
 #include <linux/lz4.h>
diff --git a/fs/bcachefs/data_update.c b/fs/bcachefs/data_update.c
index 81518f20d37d..29576c4c109d 100644
--- a/fs/bcachefs/data_update.c
+++ b/fs/bcachefs/data_update.c
@@ -9,7 +9,7 @@
 #include "ec.h"
 #include "error.h"
 #include "extents.h"
-#include "io.h"
+#include "io_write.h"
 #include "keylist.h"
 #include "move.h"
 #include "nocow_locking.h"
diff --git a/fs/bcachefs/data_update.h b/fs/bcachefs/data_update.h
index 49e9055cbb52..7ca1f98d7e94 100644
--- a/fs/bcachefs/data_update.h
+++ b/fs/bcachefs/data_update.h
@@ -4,7 +4,7 @@
 #define _BCACHEFS_DATA_UPDATE_H
 
 #include "bkey_buf.h"
-#include "io_types.h"
+#include "io_write_types.h"
 
 struct moving_context;
 
diff --git a/fs/bcachefs/debug.c b/fs/bcachefs/debug.c
index ae47e1854b80..5f3e65f9069e 100644
--- a/fs/bcachefs/debug.c
+++ b/fs/bcachefs/debug.c
@@ -19,7 +19,6 @@
 #include "extents.h"
 #include "fsck.h"
 #include "inode.h"
-#include "io.h"
 #include "super.h"
 
 #include <linux/console.h>
diff --git a/fs/bcachefs/ec.c b/fs/bcachefs/ec.c
index 67a5453a36d9..40e72b96745a 100644
--- a/fs/bcachefs/ec.c
+++ b/fs/bcachefs/ec.c
@@ -11,10 +11,11 @@
 #include "btree_update.h"
 #include "btree_write_buffer.h"
 #include "buckets.h"
+#include "checksum.h"
 #include "disk_groups.h"
 #include "ec.h"
 #include "error.h"
-#include "io.h"
+#include "io_read.h"
 #include "keylist.h"
 #include "recovery.h"
 #include "replicas.h"
diff --git a/fs/bcachefs/errcode.c b/fs/bcachefs/errcode.c
index dc906fc9176f..8d58f2cca260 100644
--- a/fs/bcachefs/errcode.c
+++ b/fs/bcachefs/errcode.c
@@ -61,3 +61,10 @@ int __bch2_err_class(int err)
 
 	return -err;
 }
+
+const char *bch2_blk_status_to_str(blk_status_t status)
+{
+	if (status == BLK_STS_REMOVED)
+		return "device removed";
+	return blk_status_to_str(status);
+}
diff --git a/fs/bcachefs/errcode.h b/fs/bcachefs/errcode.h
index f7fa87442e98..379d9d7ed333 100644
--- a/fs/bcachefs/errcode.h
+++ b/fs/bcachefs/errcode.h
@@ -249,4 +249,8 @@ static inline long bch2_err_class(long err)
 	return err < 0 ? __bch2_err_class(err) : err;
 }
 
+#define BLK_STS_REMOVED		((__force blk_status_t)128)
+
+const char *bch2_blk_status_to_str(blk_status_t);
+
 #endif /* _BCACHFES_ERRCODE_H */
diff --git a/fs/bcachefs/error.c b/fs/bcachefs/error.c
index 39009cf0c448..2a5af8872613 100644
--- a/fs/bcachefs/error.c
+++ b/fs/bcachefs/error.c
@@ -1,7 +1,6 @@
 // SPDX-License-Identifier: GPL-2.0
 #include "bcachefs.h"
 #include "error.h"
-#include "io.h"
 #include "super.h"
 
 #define FSCK_ERR_RATELIMIT_NR	10
diff --git a/fs/bcachefs/fs-io-buffered.c b/fs/bcachefs/fs-io-buffered.c
index dc22182d532f..2034d635c718 100644
--- a/fs/bcachefs/fs-io-buffered.c
+++ b/fs/bcachefs/fs-io-buffered.c
@@ -8,7 +8,8 @@
 #include "fs-io-buffered.h"
 #include "fs-io-direct.h"
 #include "fs-io-pagecache.h"
-#include "io.h"
+#include "io_read.h"
+#include "io_write.h"
 
 #include <linux/backing-dev.h>
 #include <linux/pagemap.h>
diff --git a/fs/bcachefs/fs-io-direct.c b/fs/bcachefs/fs-io-direct.c
index 2b29abd24d56..219bc1124477 100644
--- a/fs/bcachefs/fs-io-direct.c
+++ b/fs/bcachefs/fs-io-direct.c
@@ -7,7 +7,8 @@
 #include "fs-io.h"
 #include "fs-io-direct.h"
 #include "fs-io-pagecache.h"
-#include "io.h"
+#include "io_read.h"
+#include "io_write.h"
 
 #include <linux/kthread.h>
 #include <linux/pagemap.h>
diff --git a/fs/bcachefs/fs-io.c b/fs/bcachefs/fs-io.c
index ceab12fb8a8f..0b0b3b0d6c7d 100644
--- a/fs/bcachefs/fs-io.c
+++ b/fs/bcachefs/fs-io.c
@@ -3,6 +3,7 @@
 
 #include "bcachefs.h"
 #include "alloc_foreground.h"
+#include "bkey_buf.h"
 #include "btree_update.h"
 #include "buckets.h"
 #include "clock.h"
@@ -16,7 +17,7 @@
 #include "fsck.h"
 #include "inode.h"
 #include "journal.h"
-#include "io.h"
+#include "io_misc.h"
 #include "keylist.h"
 #include "quota.h"
 #include "reflink.h"
diff --git a/fs/bcachefs/fs-io.h b/fs/bcachefs/fs-io.h
index bb5b709fa8cf..bc6e8439d40b 100644
--- a/fs/bcachefs/fs-io.h
+++ b/fs/bcachefs/fs-io.h
@@ -6,7 +6,7 @@
 
 #include "buckets.h"
 #include "fs.h"
-#include "io_types.h"
+#include "io_write_types.h"
 #include "quota.h"
 
 #include <linux/uio.h>
diff --git a/fs/bcachefs/fs.c b/fs/bcachefs/fs.c
index 08f810992a1b..0648874d54f3 100644
--- a/fs/bcachefs/fs.c
+++ b/fs/bcachefs/fs.c
@@ -19,7 +19,7 @@
 #include "fs-io-pagecache.h"
 #include "fsck.h"
 #include "inode.h"
-#include "io.h"
+#include "io_read.h"
 #include "journal.h"
 #include "keylist.h"
 #include "quota.h"
diff --git a/fs/bcachefs/io.h b/fs/bcachefs/io.h
deleted file mode 100644
index 831e3f1b7e41..000000000000
--- a/fs/bcachefs/io.h
+++ /dev/null
@@ -1,202 +0,0 @@
-/* SPDX-License-Identifier: GPL-2.0 */
-#ifndef _BCACHEFS_IO_H
-#define _BCACHEFS_IO_H
-
-#include "checksum.h"
-#include "bkey_buf.h"
-#include "io_types.h"
-
-#define to_wbio(_bio)			\
-	container_of((_bio), struct bch_write_bio, bio)
-
-#define to_rbio(_bio)			\
-	container_of((_bio), struct bch_read_bio, bio)
-
-void bch2_bio_free_pages_pool(struct bch_fs *, struct bio *);
-void bch2_bio_alloc_pages_pool(struct bch_fs *, struct bio *, size_t);
-
-#ifndef CONFIG_BCACHEFS_NO_LATENCY_ACCT
-void bch2_latency_acct(struct bch_dev *, u64, int);
-#else
-static inline void bch2_latency_acct(struct bch_dev *ca, u64 submit_time, int rw) {}
-#endif
-
-void bch2_submit_wbio_replicas(struct bch_write_bio *, struct bch_fs *,
-			       enum bch_data_type, const struct bkey_i *, bool);
-
-#define BLK_STS_REMOVED		((__force blk_status_t)128)
-
-const char *bch2_blk_status_to_str(blk_status_t);
-
-#define BCH_WRITE_FLAGS()		\
-	x(ALLOC_NOWAIT)			\
-	x(CACHED)			\
-	x(DATA_ENCODED)			\
-	x(PAGES_STABLE)			\
-	x(PAGES_OWNED)			\
-	x(ONLY_SPECIFIED_DEVS)		\
-	x(WROTE_DATA_INLINE)		\
-	x(FROM_INTERNAL)		\
-	x(CHECK_ENOSPC)			\
-	x(SYNC)				\
-	x(MOVE)				\
-	x(IN_WORKER)			\
-	x(DONE)				\
-	x(IO_ERROR)			\
-	x(CONVERT_UNWRITTEN)
-
-enum __bch_write_flags {
-#define x(f)	__BCH_WRITE_##f,
-	BCH_WRITE_FLAGS()
-#undef x
-};
-
-enum bch_write_flags {
-#define x(f)	BCH_WRITE_##f = BIT(__BCH_WRITE_##f),
-	BCH_WRITE_FLAGS()
-#undef x
-};
-
-static inline struct workqueue_struct *index_update_wq(struct bch_write_op *op)
-{
-	return op->watermark == BCH_WATERMARK_copygc
-		? op->c->copygc_wq
-		: op->c->btree_update_wq;
-}
-
-int bch2_sum_sector_overwrites(struct btree_trans *, struct btree_iter *,
-			       struct bkey_i *, bool *, s64 *, s64 *);
-int bch2_extent_update(struct btree_trans *, subvol_inum,
-		       struct btree_iter *, struct bkey_i *,
-		       struct disk_reservation *, u64, s64 *, bool);
-int bch2_extent_fallocate(struct btree_trans *, subvol_inum, struct btree_iter *,
-			  unsigned, struct bch_io_opts, s64 *,
-			  struct write_point_specifier);
-
-int bch2_fpunch_at(struct btree_trans *, struct btree_iter *,
-		   subvol_inum, u64, s64 *);
-int bch2_fpunch(struct bch_fs *c, subvol_inum, u64, u64, s64 *);
-
-static inline void bch2_write_op_init(struct bch_write_op *op, struct bch_fs *c,
-				      struct bch_io_opts opts)
-{
-	op->c			= c;
-	op->end_io		= NULL;
-	op->flags		= 0;
-	op->written		= 0;
-	op->error		= 0;
-	op->csum_type		= bch2_data_checksum_type(c, opts);
-	op->compression_opt	= opts.compression;
-	op->nr_replicas		= 0;
-	op->nr_replicas_required = c->opts.data_replicas_required;
-	op->watermark		= BCH_WATERMARK_normal;
-	op->incompressible	= 0;
-	op->open_buckets.nr	= 0;
-	op->devs_have.nr	= 0;
-	op->target		= 0;
-	op->opts		= opts;
-	op->subvol		= 0;
-	op->pos			= POS_MAX;
-	op->version		= ZERO_VERSION;
-	op->write_point		= (struct write_point_specifier) { 0 };
-	op->res			= (struct disk_reservation) { 0 };
-	op->new_i_size		= U64_MAX;
-	op->i_sectors_delta	= 0;
-	op->devs_need_flush	= NULL;
-}
-
-void bch2_write(struct closure *);
-
-void bch2_write_point_do_index_updates(struct work_struct *);
-
-static inline struct bch_write_bio *wbio_init(struct bio *bio)
-{
-	struct bch_write_bio *wbio = to_wbio(bio);
-
-	memset(&wbio->wbio, 0, sizeof(wbio->wbio));
-	return wbio;
-}
-
-void bch2_write_op_to_text(struct printbuf *, struct bch_write_op *);
-
-struct bch_devs_mask;
-struct cache_promote_op;
-struct extent_ptr_decoded;
-
-int __bch2_read_indirect_extent(struct btree_trans *, unsigned *,
-				struct bkey_buf *);
-
-static inline int bch2_read_indirect_extent(struct btree_trans *trans,
-					    enum btree_id *data_btree,
-					    unsigned *offset_into_extent,
-					    struct bkey_buf *k)
-{
-	if (k->k->k.type != KEY_TYPE_reflink_p)
-		return 0;
-
-	*data_btree = BTREE_ID_reflink;
-	return __bch2_read_indirect_extent(trans, offset_into_extent, k);
-}
-
-enum bch_read_flags {
-	BCH_READ_RETRY_IF_STALE		= 1 << 0,
-	BCH_READ_MAY_PROMOTE		= 1 << 1,
-	BCH_READ_USER_MAPPED		= 1 << 2,
-	BCH_READ_NODECODE		= 1 << 3,
-	BCH_READ_LAST_FRAGMENT		= 1 << 4,
-
-	/* internal: */
-	BCH_READ_MUST_BOUNCE		= 1 << 5,
-	BCH_READ_MUST_CLONE		= 1 << 6,
-	BCH_READ_IN_RETRY		= 1 << 7,
-};
-
-int __bch2_read_extent(struct btree_trans *, struct bch_read_bio *,
-		       struct bvec_iter, struct bpos, enum btree_id,
-		       struct bkey_s_c, unsigned,
-		       struct bch_io_failures *, unsigned);
-
-static inline void bch2_read_extent(struct btree_trans *trans,
-			struct bch_read_bio *rbio, struct bpos read_pos,
-			enum btree_id data_btree, struct bkey_s_c k,
-			unsigned offset_into_extent, unsigned flags)
-{
-	__bch2_read_extent(trans, rbio, rbio->bio.bi_iter, read_pos,
-			   data_btree, k, offset_into_extent, NULL, flags);
-}
-
-void __bch2_read(struct bch_fs *, struct bch_read_bio *, struct bvec_iter,
-		 subvol_inum, struct bch_io_failures *, unsigned flags);
-
-static inline void bch2_read(struct bch_fs *c, struct bch_read_bio *rbio,
-			     subvol_inum inum)
-{
-	struct bch_io_failures failed = { .nr = 0 };
-
-	BUG_ON(rbio->_state);
-
-	rbio->c = c;
-	rbio->start_time = local_clock();
-	rbio->subvol = inum.subvol;
-
-	__bch2_read(c, rbio, rbio->bio.bi_iter, inum, &failed,
-		    BCH_READ_RETRY_IF_STALE|
-		    BCH_READ_MAY_PROMOTE|
-		    BCH_READ_USER_MAPPED);
-}
-
-static inline struct bch_read_bio *rbio_init(struct bio *bio,
-					     struct bch_io_opts opts)
-{
-	struct bch_read_bio *rbio = to_rbio(bio);
-
-	rbio->_state	= 0;
-	rbio->promote	= NULL;
-	rbio->opts	= opts;
-	return rbio;
-}
-
-void bch2_fs_io_exit(struct bch_fs *);
-int bch2_fs_io_init(struct bch_fs *);
-
-#endif /* _BCACHEFS_IO_H */
diff --git a/fs/bcachefs/io_misc.c b/fs/bcachefs/io_misc.c
new file mode 100644
index 000000000000..c04e5dacfc8d
--- /dev/null
+++ b/fs/bcachefs/io_misc.c
@@ -0,0 +1,215 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * io_misc.c - fallocate, fpunch, truncate:
+ */
+
+#include "bcachefs.h"
+#include "alloc_foreground.h"
+#include "bkey_buf.h"
+#include "btree_update.h"
+#include "buckets.h"
+#include "clock.h"
+#include "extents.h"
+#include "io_misc.h"
+#include "io_write.h"
+#include "subvolume.h"
+
+/* Overwrites whatever was present with zeroes: */
+int bch2_extent_fallocate(struct btree_trans *trans,
+			  subvol_inum inum,
+			  struct btree_iter *iter,
+			  unsigned sectors,
+			  struct bch_io_opts opts,
+			  s64 *i_sectors_delta,
+			  struct write_point_specifier write_point)
+{
+	struct bch_fs *c = trans->c;
+	struct disk_reservation disk_res = { 0 };
+	struct closure cl;
+	struct open_buckets open_buckets = { 0 };
+	struct bkey_s_c k;
+	struct bkey_buf old, new;
+	unsigned sectors_allocated = 0;
+	bool have_reservation = false;
+	bool unwritten = opts.nocow &&
+	    c->sb.version >= bcachefs_metadata_version_unwritten_extents;
+	int ret;
+
+	bch2_bkey_buf_init(&old);
+	bch2_bkey_buf_init(&new);
+	closure_init_stack(&cl);
+
+	k = bch2_btree_iter_peek_slot(iter);
+	ret = bkey_err(k);
+	if (ret)
+		return ret;
+
+	sectors = min_t(u64, sectors, k.k->p.offset - iter->pos.offset);
+
+	if (!have_reservation) {
+		unsigned new_replicas =
+			max(0, (int) opts.data_replicas -
+			    (int) bch2_bkey_nr_ptrs_fully_allocated(k));
+		/*
+		 * Get a disk reservation before (in the nocow case) calling
+		 * into the allocator:
+		 */
+		ret = bch2_disk_reservation_get(c, &disk_res, sectors, new_replicas, 0);
+		if (unlikely(ret))
+			goto err;
+
+		bch2_bkey_buf_reassemble(&old, c, k);
+	}
+
+	if (have_reservation) {
+		if (!bch2_extents_match(k, bkey_i_to_s_c(old.k)))
+			goto err;
+
+		bch2_key_resize(&new.k->k, sectors);
+	} else if (!unwritten) {
+		struct bkey_i_reservation *reservation;
+
+		bch2_bkey_buf_realloc(&new, c, sizeof(*reservation) / sizeof(u64));
+		reservation = bkey_reservation_init(new.k);
+		reservation->k.p = iter->pos;
+		bch2_key_resize(&reservation->k, sectors);
+		reservation->v.nr_replicas = opts.data_replicas;
+	} else {
+		struct bkey_i_extent *e;
+		struct bch_devs_list devs_have;
+		struct write_point *wp;
+		struct bch_extent_ptr *ptr;
+
+		devs_have.nr = 0;
+
+		bch2_bkey_buf_realloc(&new, c, BKEY_EXTENT_U64s_MAX);
+
+		e = bkey_extent_init(new.k);
+		e->k.p = iter->pos;
+
+		ret = bch2_alloc_sectors_start_trans(trans,
+				opts.foreground_target,
+				false,
+				write_point,
+				&devs_have,
+				opts.data_replicas,
+				opts.data_replicas,
+				BCH_WATERMARK_normal, 0, &cl, &wp);
+		if (bch2_err_matches(ret, BCH_ERR_operation_blocked))
+			ret = -BCH_ERR_transaction_restart_nested;
+		if (ret)
+			goto err;
+
+		sectors = min(sectors, wp->sectors_free);
+		sectors_allocated = sectors;
+
+		bch2_key_resize(&e->k, sectors);
+
+		bch2_open_bucket_get(c, wp, &open_buckets);
+		bch2_alloc_sectors_append_ptrs(c, wp, &e->k_i, sectors, false);
+		bch2_alloc_sectors_done(c, wp);
+
+		extent_for_each_ptr(extent_i_to_s(e), ptr)
+			ptr->unwritten = true;
+	}
+
+	have_reservation = true;
+
+	ret = bch2_extent_update(trans, inum, iter, new.k, &disk_res,
+				 0, i_sectors_delta, true);
+err:
+	if (!ret && sectors_allocated)
+		bch2_increment_clock(c, sectors_allocated, WRITE);
+
+	bch2_open_buckets_put(c, &open_buckets);
+	bch2_disk_reservation_put(c, &disk_res);
+	bch2_bkey_buf_exit(&new, c);
+	bch2_bkey_buf_exit(&old, c);
+
+	if (closure_nr_remaining(&cl) != 1) {
+		bch2_trans_unlock(trans);
+		closure_sync(&cl);
+	}
+
+	return ret;
+}
+
+/*
+ * Returns -BCH_ERR_transacton_restart if we had to drop locks:
+ */
+int bch2_fpunch_at(struct btree_trans *trans, struct btree_iter *iter,
+		   subvol_inum inum, u64 end,
+		   s64 *i_sectors_delta)
+{
+	struct bch_fs *c	= trans->c;
+	unsigned max_sectors	= KEY_SIZE_MAX & (~0 << c->block_bits);
+	struct bpos end_pos = POS(inum.inum, end);
+	struct bkey_s_c k;
+	int ret = 0, ret2 = 0;
+	u32 snapshot;
+
+	while (!ret ||
+	       bch2_err_matches(ret, BCH_ERR_transaction_restart)) {
+		struct disk_reservation disk_res =
+			bch2_disk_reservation_init(c, 0);
+		struct bkey_i delete;
+
+		if (ret)
+			ret2 = ret;
+
+		bch2_trans_begin(trans);
+
+		ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
+		if (ret)
+			continue;
+
+		bch2_btree_iter_set_snapshot(iter, snapshot);
+
+		/*
+		 * peek_upto() doesn't have ideal semantics for extents:
+		 */
+		k = bch2_btree_iter_peek_upto(iter, end_pos);
+		if (!k.k)
+			break;
+
+		ret = bkey_err(k);
+		if (ret)
+			continue;
+
+		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_pos, &delete);
+
+		ret = bch2_extent_update(trans, inum, iter, &delete,
+				&disk_res, 0, i_sectors_delta, false);
+		bch2_disk_reservation_put(c, &disk_res);
+	}
+
+	return ret ?: ret2;
+}
+
+int bch2_fpunch(struct bch_fs *c, subvol_inum inum, u64 start, u64 end,
+		s64 *i_sectors_delta)
+{
+	struct btree_trans trans;
+	struct btree_iter iter;
+	int ret;
+
+	bch2_trans_init(&trans, c, BTREE_ITER_MAX, 1024);
+	bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents,
+			     POS(inum.inum, start),
+			     BTREE_ITER_INTENT);
+
+	ret = bch2_fpunch_at(&trans, &iter, inum, end, i_sectors_delta);
+
+	bch2_trans_iter_exit(&trans, &iter);
+	bch2_trans_exit(&trans);
+
+	if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
+		ret = 0;
+
+	return ret;
+}
diff --git a/fs/bcachefs/io_misc.h b/fs/bcachefs/io_misc.h
new file mode 100644
index 000000000000..46e9ce3251d6
--- /dev/null
+++ b/fs/bcachefs/io_misc.h
@@ -0,0 +1,12 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _BCACHEFS_IO_MISC_H
+#define _BCACHEFS_IO_MISC_H
+
+int bch2_extent_fallocate(struct btree_trans *, subvol_inum, struct btree_iter *,
+			  unsigned, struct bch_io_opts, s64 *,
+			  struct write_point_specifier);
+int bch2_fpunch_at(struct btree_trans *, struct btree_iter *,
+		   subvol_inum, u64, s64 *);
+int bch2_fpunch(struct bch_fs *c, subvol_inum, u64, u64, s64 *);
+
+#endif /* _BCACHEFS_IO_MISC_H */
diff --git a/fs/bcachefs/io_read.c b/fs/bcachefs/io_read.c
new file mode 100644
index 000000000000..cd62bf730396
--- /dev/null
+++ b/fs/bcachefs/io_read.c
@@ -0,0 +1,1207 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Some low level IO code, and hacks for various block layer limitations
+ *
+ * Copyright 2010, 2011 Kent Overstreet <kent.overstreet@gmail.com>
+ * Copyright 2012 Google, Inc.
+ */
+
+#include "bcachefs.h"
+#include "alloc_background.h"
+#include "alloc_foreground.h"
+#include "btree_update.h"
+#include "buckets.h"
+#include "checksum.h"
+#include "clock.h"
+#include "compress.h"
+#include "data_update.h"
+#include "disk_groups.h"
+#include "ec.h"
+#include "error.h"
+#include "io_read.h"
+#include "io_misc.h"
+#include "io_write.h"
+#include "subvolume.h"
+#include "trace.h"
+
+#ifndef CONFIG_BCACHEFS_NO_LATENCY_ACCT
+
+static bool bch2_target_congested(struct bch_fs *c, u16 target)
+{
+	const struct bch_devs_mask *devs;
+	unsigned d, nr = 0, total = 0;
+	u64 now = local_clock(), last;
+	s64 congested;
+	struct bch_dev *ca;
+
+	if (!target)
+		return false;
+
+	rcu_read_lock();
+	devs = bch2_target_to_mask(c, target) ?:
+		&c->rw_devs[BCH_DATA_user];
+
+	for_each_set_bit(d, devs->d, BCH_SB_MEMBERS_MAX) {
+		ca = rcu_dereference(c->devs[d]);
+		if (!ca)
+			continue;
+
+		congested = atomic_read(&ca->congested);
+		last = READ_ONCE(ca->congested_last);
+		if (time_after64(now, last))
+			congested -= (now - last) >> 12;
+
+		total += max(congested, 0LL);
+		nr++;
+	}
+	rcu_read_unlock();
+
+	return bch2_rand_range(nr * CONGESTED_MAX) < total;
+}
+
+#else
+
+static bool bch2_target_congested(struct bch_fs *c, u16 target)
+{
+	return false;
+}
+
+#endif
+
+/* Cache promotion on read */
+
+struct promote_op {
+	struct rcu_head		rcu;
+	u64			start_time;
+
+	struct rhash_head	hash;
+	struct bpos		pos;
+
+	struct data_update	write;
+	struct bio_vec		bi_inline_vecs[0]; /* must be last */
+};
+
+static const struct rhashtable_params bch_promote_params = {
+	.head_offset	= offsetof(struct promote_op, hash),
+	.key_offset	= offsetof(struct promote_op, pos),
+	.key_len	= sizeof(struct bpos),
+};
+
+static inline bool should_promote(struct bch_fs *c, struct bkey_s_c k,
+				  struct bpos pos,
+				  struct bch_io_opts opts,
+				  unsigned flags)
+{
+	if (!(flags & BCH_READ_MAY_PROMOTE))
+		return false;
+
+	if (!opts.promote_target)
+		return false;
+
+	if (bch2_bkey_has_target(c, k, opts.promote_target))
+		return false;
+
+	if (bkey_extent_is_unwritten(k))
+		return false;
+
+	if (bch2_target_congested(c, opts.promote_target)) {
+		/* XXX trace this */
+		return false;
+	}
+
+	if (rhashtable_lookup_fast(&c->promote_table, &pos,
+				   bch_promote_params))
+		return false;
+
+	return true;
+}
+
+static void promote_free(struct bch_fs *c, struct promote_op *op)
+{
+	int ret;
+
+	bch2_data_update_exit(&op->write);
+
+	ret = rhashtable_remove_fast(&c->promote_table, &op->hash,
+				     bch_promote_params);
+	BUG_ON(ret);
+	bch2_write_ref_put(c, BCH_WRITE_REF_promote);
+	kfree_rcu(op, rcu);
+}
+
+static void promote_done(struct bch_write_op *wop)
+{
+	struct promote_op *op =
+		container_of(wop, struct promote_op, write.op);
+	struct bch_fs *c = op->write.op.c;
+
+	bch2_time_stats_update(&c->times[BCH_TIME_data_promote],
+			       op->start_time);
+	promote_free(c, op);
+}
+
+static void promote_start(struct promote_op *op, struct bch_read_bio *rbio)
+{
+	struct bio *bio = &op->write.op.wbio.bio;
+
+	trace_and_count(op->write.op.c, read_promote, &rbio->bio);
+
+	/* we now own pages: */
+	BUG_ON(!rbio->bounce);
+	BUG_ON(rbio->bio.bi_vcnt > bio->bi_max_vecs);
+
+	memcpy(bio->bi_io_vec, rbio->bio.bi_io_vec,
+	       sizeof(struct bio_vec) * rbio->bio.bi_vcnt);
+	swap(bio->bi_vcnt, rbio->bio.bi_vcnt);
+
+	bch2_data_update_read_done(&op->write, rbio->pick.crc);
+}
+
+static struct promote_op *__promote_alloc(struct btree_trans *trans,
+					  enum btree_id btree_id,
+					  struct bkey_s_c k,
+					  struct bpos pos,
+					  struct extent_ptr_decoded *pick,
+					  struct bch_io_opts opts,
+					  unsigned sectors,
+					  struct bch_read_bio **rbio)
+{
+	struct bch_fs *c = trans->c;
+	struct promote_op *op = NULL;
+	struct bio *bio;
+	unsigned pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
+	int ret;
+
+	if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_promote))
+		return NULL;
+
+	op = kzalloc(sizeof(*op) + sizeof(struct bio_vec) * pages, GFP_NOFS);
+	if (!op)
+		goto err;
+
+	op->start_time = local_clock();
+	op->pos = pos;
+
+	/*
+	 * We don't use the mempool here because extents that aren't
+	 * checksummed or compressed can be too big for the mempool:
+	 */
+	*rbio = kzalloc(sizeof(struct bch_read_bio) +
+			sizeof(struct bio_vec) * pages,
+			GFP_NOFS);
+	if (!*rbio)
+		goto err;
+
+	rbio_init(&(*rbio)->bio, opts);
+	bio_init(&(*rbio)->bio, NULL, (*rbio)->bio.bi_inline_vecs, pages, 0);
+
+	if (bch2_bio_alloc_pages(&(*rbio)->bio, sectors << 9,
+				 GFP_NOFS))
+		goto err;
+
+	(*rbio)->bounce		= true;
+	(*rbio)->split		= true;
+	(*rbio)->kmalloc	= true;
+
+	if (rhashtable_lookup_insert_fast(&c->promote_table, &op->hash,
+					  bch_promote_params))
+		goto err;
+
+	bio = &op->write.op.wbio.bio;
+	bio_init(bio, NULL, bio->bi_inline_vecs, pages, 0);
+
+	ret = bch2_data_update_init(trans, NULL, &op->write,
+			writepoint_hashed((unsigned long) current),
+			opts,
+			(struct data_update_opts) {
+				.target		= opts.promote_target,
+				.extra_replicas	= 1,
+				.write_flags	= BCH_WRITE_ALLOC_NOWAIT|BCH_WRITE_CACHED,
+			},
+			btree_id, k);
+	/*
+	 * possible errors: -BCH_ERR_nocow_lock_blocked,
+	 * -BCH_ERR_ENOSPC_disk_reservation:
+	 */
+	if (ret) {
+		ret = rhashtable_remove_fast(&c->promote_table, &op->hash,
+					bch_promote_params);
+		BUG_ON(ret);
+		goto err;
+	}
+
+	op->write.op.end_io = promote_done;
+
+	return op;
+err:
+	if (*rbio)
+		bio_free_pages(&(*rbio)->bio);
+	kfree(*rbio);
+	*rbio = NULL;
+	kfree(op);
+	bch2_write_ref_put(c, BCH_WRITE_REF_promote);
+	return NULL;
+}
+
+noinline
+static struct promote_op *promote_alloc(struct btree_trans *trans,
+					struct bvec_iter iter,
+					struct bkey_s_c k,
+					struct extent_ptr_decoded *pick,
+					struct bch_io_opts opts,
+					unsigned flags,
+					struct bch_read_bio **rbio,
+					bool *bounce,
+					bool *read_full)
+{
+	struct bch_fs *c = trans->c;
+	bool promote_full = *read_full || READ_ONCE(c->promote_whole_extents);
+	/* data might have to be decompressed in the write path: */
+	unsigned sectors = promote_full
+		? max(pick->crc.compressed_size, pick->crc.live_size)
+		: bvec_iter_sectors(iter);
+	struct bpos pos = promote_full
+		? bkey_start_pos(k.k)
+		: POS(k.k->p.inode, iter.bi_sector);
+	struct promote_op *promote;
+
+	if (!should_promote(c, k, pos, opts, flags))
+		return NULL;
+
+	promote = __promote_alloc(trans,
+				  k.k->type == KEY_TYPE_reflink_v
+				  ? BTREE_ID_reflink
+				  : BTREE_ID_extents,
+				  k, pos, pick, opts, sectors, rbio);
+	if (!promote)
+		return NULL;
+
+	*bounce		= true;
+	*read_full	= promote_full;
+	return promote;
+}
+
+/* Read */
+
+#define READ_RETRY_AVOID	1
+#define READ_RETRY		2
+#define READ_ERR		3
+
+enum rbio_context {
+	RBIO_CONTEXT_NULL,
+	RBIO_CONTEXT_HIGHPRI,
+	RBIO_CONTEXT_UNBOUND,
+};
+
+static inline struct bch_read_bio *
+bch2_rbio_parent(struct bch_read_bio *rbio)
+{
+	return rbio->split ? rbio->parent : rbio;
+}
+
+__always_inline
+static void bch2_rbio_punt(struct bch_read_bio *rbio, work_func_t fn,
+			   enum rbio_context context,
+			   struct workqueue_struct *wq)
+{
+	if (context <= rbio->context) {
+		fn(&rbio->work);
+	} else {
+		rbio->work.func		= fn;
+		rbio->context		= context;
+		queue_work(wq, &rbio->work);
+	}
+}
+
+static inline struct bch_read_bio *bch2_rbio_free(struct bch_read_bio *rbio)
+{
+	BUG_ON(rbio->bounce && !rbio->split);
+
+	if (rbio->promote)
+		promote_free(rbio->c, rbio->promote);
+	rbio->promote = NULL;
+
+	if (rbio->bounce)
+		bch2_bio_free_pages_pool(rbio->c, &rbio->bio);
+
+	if (rbio->split) {
+		struct bch_read_bio *parent = rbio->parent;
+
+		if (rbio->kmalloc)
+			kfree(rbio);
+		else
+			bio_put(&rbio->bio);
+
+		rbio = parent;
+	}
+
+	return rbio;
+}
+
+/*
+ * Only called on a top level bch_read_bio to complete an entire read request,
+ * not a split:
+ */
+static void bch2_rbio_done(struct bch_read_bio *rbio)
+{
+	if (rbio->start_time)
+		bch2_time_stats_update(&rbio->c->times[BCH_TIME_data_read],
+				       rbio->start_time);
+	bio_endio(&rbio->bio);
+}
+
+static void bch2_read_retry_nodecode(struct bch_fs *c, struct bch_read_bio *rbio,
+				     struct bvec_iter bvec_iter,
+				     struct bch_io_failures *failed,
+				     unsigned flags)
+{
+	struct btree_trans trans;
+	struct btree_iter iter;
+	struct bkey_buf sk;
+	struct bkey_s_c k;
+	int ret;
+
+	flags &= ~BCH_READ_LAST_FRAGMENT;
+	flags |= BCH_READ_MUST_CLONE;
+
+	bch2_bkey_buf_init(&sk);
+	bch2_trans_init(&trans, c, 0, 0);
+
+	bch2_trans_iter_init(&trans, &iter, rbio->data_btree,
+			     rbio->read_pos, BTREE_ITER_SLOTS);
+retry:
+	rbio->bio.bi_status = 0;
+
+	k = bch2_btree_iter_peek_slot(&iter);
+	if (bkey_err(k))
+		goto err;
+
+	bch2_bkey_buf_reassemble(&sk, c, k);
+	k = bkey_i_to_s_c(sk.k);
+	bch2_trans_unlock(&trans);
+
+	if (!bch2_bkey_matches_ptr(c, k,
+				   rbio->pick.ptr,
+				   rbio->data_pos.offset -
+				   rbio->pick.crc.offset)) {
+		/* extent we wanted to read no longer exists: */
+		rbio->hole = true;
+		goto out;
+	}
+
+	ret = __bch2_read_extent(&trans, rbio, bvec_iter,
+				 rbio->read_pos,
+				 rbio->data_btree,
+				 k, 0, failed, flags);
+	if (ret == READ_RETRY)
+		goto retry;
+	if (ret)
+		goto err;
+out:
+	bch2_rbio_done(rbio);
+	bch2_trans_iter_exit(&trans, &iter);
+	bch2_trans_exit(&trans);
+	bch2_bkey_buf_exit(&sk, c);
+	return;
+err:
+	rbio->bio.bi_status = BLK_STS_IOERR;
+	goto out;
+}
+
+static void bch2_rbio_retry(struct work_struct *work)
+{
+	struct bch_read_bio *rbio =
+		container_of(work, struct bch_read_bio, work);
+	struct bch_fs *c	= rbio->c;
+	struct bvec_iter iter	= rbio->bvec_iter;
+	unsigned flags		= rbio->flags;
+	subvol_inum inum = {
+		.subvol = rbio->subvol,
+		.inum	= rbio->read_pos.inode,
+	};
+	struct bch_io_failures failed = { .nr = 0 };
+
+	trace_and_count(c, read_retry, &rbio->bio);
+
+	if (rbio->retry == READ_RETRY_AVOID)
+		bch2_mark_io_failure(&failed, &rbio->pick);
+
+	rbio->bio.bi_status = 0;
+
+	rbio = bch2_rbio_free(rbio);
+
+	flags |= BCH_READ_IN_RETRY;
+	flags &= ~BCH_READ_MAY_PROMOTE;
+
+	if (flags & BCH_READ_NODECODE) {
+		bch2_read_retry_nodecode(c, rbio, iter, &failed, flags);
+	} else {
+		flags &= ~BCH_READ_LAST_FRAGMENT;
+		flags |= BCH_READ_MUST_CLONE;
+
+		__bch2_read(c, rbio, iter, inum, &failed, flags);
+	}
+}
+
+static void bch2_rbio_error(struct bch_read_bio *rbio, int retry,
+			    blk_status_t error)
+{
+	rbio->retry = retry;
+
+	if (rbio->flags & BCH_READ_IN_RETRY)
+		return;
+
+	if (retry == READ_ERR) {
+		rbio = bch2_rbio_free(rbio);
+
+		rbio->bio.bi_status = error;
+		bch2_rbio_done(rbio);
+	} else {
+		bch2_rbio_punt(rbio, bch2_rbio_retry,
+			       RBIO_CONTEXT_UNBOUND, system_unbound_wq);
+	}
+}
+
+static int __bch2_rbio_narrow_crcs(struct btree_trans *trans,
+				   struct bch_read_bio *rbio)
+{
+	struct bch_fs *c = rbio->c;
+	u64 data_offset = rbio->data_pos.offset - rbio->pick.crc.offset;
+	struct bch_extent_crc_unpacked new_crc;
+	struct btree_iter iter;
+	struct bkey_i *new;
+	struct bkey_s_c k;
+	int ret = 0;
+
+	if (crc_is_compressed(rbio->pick.crc))
+		return 0;
+
+	k = bch2_bkey_get_iter(trans, &iter, rbio->data_btree, rbio->data_pos,
+			       BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
+	if ((ret = bkey_err(k)))
+		goto out;
+
+	if (bversion_cmp(k.k->version, rbio->version) ||
+	    !bch2_bkey_matches_ptr(c, k, rbio->pick.ptr, data_offset))
+		goto out;
+
+	/* Extent was merged? */
+	if (bkey_start_offset(k.k) < data_offset ||
+	    k.k->p.offset > data_offset + rbio->pick.crc.uncompressed_size)
+		goto out;
+
+	if (bch2_rechecksum_bio(c, &rbio->bio, rbio->version,
+			rbio->pick.crc, NULL, &new_crc,
+			bkey_start_offset(k.k) - data_offset, k.k->size,
+			rbio->pick.crc.csum_type)) {
+		bch_err(c, "error verifying existing checksum while narrowing checksum (memory corruption?)");
+		ret = 0;
+		goto out;
+	}
+
+	/*
+	 * going to be temporarily appending another checksum entry:
+	 */
+	new = bch2_trans_kmalloc(trans, bkey_bytes(k.k) +
+				 sizeof(struct bch_extent_crc128));
+	if ((ret = PTR_ERR_OR_ZERO(new)))
+		goto out;
+
+	bkey_reassemble(new, k);
+
+	if (!bch2_bkey_narrow_crcs(new, new_crc))
+		goto out;
+
+	ret = bch2_trans_update(trans, &iter, new,
+				BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
+out:
+	bch2_trans_iter_exit(trans, &iter);
+	return ret;
+}
+
+static noinline void bch2_rbio_narrow_crcs(struct bch_read_bio *rbio)
+{
+	bch2_trans_do(rbio->c, NULL, NULL, BTREE_INSERT_NOFAIL,
+		      __bch2_rbio_narrow_crcs(&trans, rbio));
+}
+
+/* Inner part that may run in process context */
+static void __bch2_read_endio(struct work_struct *work)
+{
+	struct bch_read_bio *rbio =
+		container_of(work, struct bch_read_bio, work);
+	struct bch_fs *c	= rbio->c;
+	struct bch_dev *ca	= bch_dev_bkey_exists(c, rbio->pick.ptr.dev);
+	struct bio *src		= &rbio->bio;
+	struct bio *dst		= &bch2_rbio_parent(rbio)->bio;
+	struct bvec_iter dst_iter = rbio->bvec_iter;
+	struct bch_extent_crc_unpacked crc = rbio->pick.crc;
+	struct nonce nonce = extent_nonce(rbio->version, crc);
+	unsigned nofs_flags;
+	struct bch_csum csum;
+	int ret;
+
+	nofs_flags = memalloc_nofs_save();
+
+	/* Reset iterator for checksumming and copying bounced data: */
+	if (rbio->bounce) {
+		src->bi_iter.bi_size		= crc.compressed_size << 9;
+		src->bi_iter.bi_idx		= 0;
+		src->bi_iter.bi_bvec_done	= 0;
+	} else {
+		src->bi_iter			= rbio->bvec_iter;
+	}
+
+	csum = bch2_checksum_bio(c, crc.csum_type, nonce, src);
+	if (bch2_crc_cmp(csum, rbio->pick.crc.csum) && !c->opts.no_data_io)
+		goto csum_err;
+
+	/*
+	 * XXX
+	 * We need to rework the narrow_crcs path to deliver the read completion
+	 * first, and then punt to a different workqueue, otherwise we're
+	 * holding up reads while doing btree updates which is bad for memory
+	 * reclaim.
+	 */
+	if (unlikely(rbio->narrow_crcs))
+		bch2_rbio_narrow_crcs(rbio);
+
+	if (rbio->flags & BCH_READ_NODECODE)
+		goto nodecode;
+
+	/* Adjust crc to point to subset of data we want: */
+	crc.offset     += rbio->offset_into_extent;
+	crc.live_size	= bvec_iter_sectors(rbio->bvec_iter);
+
+	if (crc_is_compressed(crc)) {
+		ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src);
+		if (ret)
+			goto decrypt_err;
+
+		if (bch2_bio_uncompress(c, src, dst, dst_iter, crc) &&
+		    !c->opts.no_data_io)
+			goto decompression_err;
+	} else {
+		/* don't need to decrypt the entire bio: */
+		nonce = nonce_add(nonce, crc.offset << 9);
+		bio_advance(src, crc.offset << 9);
+
+		BUG_ON(src->bi_iter.bi_size < dst_iter.bi_size);
+		src->bi_iter.bi_size = dst_iter.bi_size;
+
+		ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src);
+		if (ret)
+			goto decrypt_err;
+
+		if (rbio->bounce) {
+			struct bvec_iter src_iter = src->bi_iter;
+
+			bio_copy_data_iter(dst, &dst_iter, src, &src_iter);
+		}
+	}
+
+	if (rbio->promote) {
+		/*
+		 * Re encrypt data we decrypted, so it's consistent with
+		 * rbio->crc:
+		 */
+		ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src);
+		if (ret)
+			goto decrypt_err;
+
+		promote_start(rbio->promote, rbio);
+		rbio->promote = NULL;
+	}
+nodecode:
+	if (likely(!(rbio->flags & BCH_READ_IN_RETRY))) {
+		rbio = bch2_rbio_free(rbio);
+		bch2_rbio_done(rbio);
+	}
+out:
+	memalloc_nofs_restore(nofs_flags);
+	return;
+csum_err:
+	/*
+	 * Checksum error: if the bio wasn't bounced, we may have been
+	 * reading into buffers owned by userspace (that userspace can
+	 * scribble over) - retry the read, bouncing it this time:
+	 */
+	if (!rbio->bounce && (rbio->flags & BCH_READ_USER_MAPPED)) {
+		rbio->flags |= BCH_READ_MUST_BOUNCE;
+		bch2_rbio_error(rbio, READ_RETRY, BLK_STS_IOERR);
+		goto out;
+	}
+
+	bch_err_inum_offset_ratelimited(ca,
+		rbio->read_pos.inode,
+		rbio->read_pos.offset << 9,
+		"data checksum error: expected %0llx:%0llx got %0llx:%0llx (type %s)",
+		rbio->pick.crc.csum.hi, rbio->pick.crc.csum.lo,
+		csum.hi, csum.lo, bch2_csum_types[crc.csum_type]);
+	bch2_io_error(ca);
+	bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
+	goto out;
+decompression_err:
+	bch_err_inum_offset_ratelimited(c, rbio->read_pos.inode,
+					rbio->read_pos.offset << 9,
+					"decompression error");
+	bch2_rbio_error(rbio, READ_ERR, BLK_STS_IOERR);
+	goto out;
+decrypt_err:
+	bch_err_inum_offset_ratelimited(c, rbio->read_pos.inode,
+					rbio->read_pos.offset << 9,
+					"decrypt error");
+	bch2_rbio_error(rbio, READ_ERR, BLK_STS_IOERR);
+	goto out;
+}
+
+static void bch2_read_endio(struct bio *bio)
+{
+	struct bch_read_bio *rbio =
+		container_of(bio, struct bch_read_bio, bio);
+	struct bch_fs *c	= rbio->c;
+	struct bch_dev *ca	= bch_dev_bkey_exists(c, rbio->pick.ptr.dev);
+	struct workqueue_struct *wq = NULL;
+	enum rbio_context context = RBIO_CONTEXT_NULL;
+
+	if (rbio->have_ioref) {
+		bch2_latency_acct(ca, rbio->submit_time, READ);
+		percpu_ref_put(&ca->io_ref);
+	}
+
+	if (!rbio->split)
+		rbio->bio.bi_end_io = rbio->end_io;
+
+	if (bch2_dev_inum_io_err_on(bio->bi_status, ca,
+				    rbio->read_pos.inode,
+				    rbio->read_pos.offset,
+				    "data read error: %s",
+			       bch2_blk_status_to_str(bio->bi_status))) {
+		bch2_rbio_error(rbio, READ_RETRY_AVOID, bio->bi_status);
+		return;
+	}
+
+	if (((rbio->flags & BCH_READ_RETRY_IF_STALE) && race_fault()) ||
+	    ptr_stale(ca, &rbio->pick.ptr)) {
+		trace_and_count(c, read_reuse_race, &rbio->bio);
+
+		if (rbio->flags & BCH_READ_RETRY_IF_STALE)
+			bch2_rbio_error(rbio, READ_RETRY, BLK_STS_AGAIN);
+		else
+			bch2_rbio_error(rbio, READ_ERR, BLK_STS_AGAIN);
+		return;
+	}
+
+	if (rbio->narrow_crcs ||
+	    rbio->promote ||
+	    crc_is_compressed(rbio->pick.crc) ||
+	    bch2_csum_type_is_encryption(rbio->pick.crc.csum_type))
+		context = RBIO_CONTEXT_UNBOUND,	wq = system_unbound_wq;
+	else if (rbio->pick.crc.csum_type)
+		context = RBIO_CONTEXT_HIGHPRI,	wq = system_highpri_wq;
+
+	bch2_rbio_punt(rbio, __bch2_read_endio, context, wq);
+}
+
+int __bch2_read_indirect_extent(struct btree_trans *trans,
+				unsigned *offset_into_extent,
+				struct bkey_buf *orig_k)
+{
+	struct btree_iter iter;
+	struct bkey_s_c k;
+	u64 reflink_offset;
+	int ret;
+
+	reflink_offset = le64_to_cpu(bkey_i_to_reflink_p(orig_k->k)->v.idx) +
+		*offset_into_extent;
+
+	k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_reflink,
+			       POS(0, reflink_offset), 0);
+	ret = bkey_err(k);
+	if (ret)
+		goto err;
+
+	if (k.k->type != KEY_TYPE_reflink_v &&
+	    k.k->type != KEY_TYPE_indirect_inline_data) {
+		bch_err_inum_offset_ratelimited(trans->c,
+			orig_k->k->k.p.inode,
+			orig_k->k->k.p.offset << 9,
+			"%llu len %u points to nonexistent indirect extent %llu",
+			orig_k->k->k.p.offset,
+			orig_k->k->k.size,
+			reflink_offset);
+		bch2_inconsistent_error(trans->c);
+		ret = -EIO;
+		goto err;
+	}
+
+	*offset_into_extent = iter.pos.offset - bkey_start_offset(k.k);
+	bch2_bkey_buf_reassemble(orig_k, trans->c, k);
+err:
+	bch2_trans_iter_exit(trans, &iter);
+	return ret;
+}
+
+static noinline void read_from_stale_dirty_pointer(struct btree_trans *trans,
+						   struct bkey_s_c k,
+						   struct bch_extent_ptr ptr)
+{
+	struct bch_fs *c = trans->c;
+	struct bch_dev *ca = bch_dev_bkey_exists(c, ptr.dev);
+	struct btree_iter iter;
+	struct printbuf buf = PRINTBUF;
+	int ret;
+
+	bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
+			     PTR_BUCKET_POS(c, &ptr),
+			     BTREE_ITER_CACHED);
+
+	prt_printf(&buf, "Attempting to read from stale dirty pointer:");
+	printbuf_indent_add(&buf, 2);
+	prt_newline(&buf);
+
+	bch2_bkey_val_to_text(&buf, c, k);
+	prt_newline(&buf);
+
+	prt_printf(&buf, "memory gen: %u", *bucket_gen(ca, iter.pos.offset));
+
+	ret = lockrestart_do(trans, bkey_err(k = bch2_btree_iter_peek_slot(&iter)));
+	if (!ret) {
+		prt_newline(&buf);
+		bch2_bkey_val_to_text(&buf, c, k);
+	}
+
+	bch2_fs_inconsistent(c, "%s", buf.buf);
+
+	bch2_trans_iter_exit(trans, &iter);
+	printbuf_exit(&buf);
+}
+
+int __bch2_read_extent(struct btree_trans *trans, struct bch_read_bio *orig,
+		       struct bvec_iter iter, struct bpos read_pos,
+		       enum btree_id data_btree, struct bkey_s_c k,
+		       unsigned offset_into_extent,
+		       struct bch_io_failures *failed, unsigned flags)
+{
+	struct bch_fs *c = trans->c;
+	struct extent_ptr_decoded pick;
+	struct bch_read_bio *rbio = NULL;
+	struct bch_dev *ca = NULL;
+	struct promote_op *promote = NULL;
+	bool bounce = false, read_full = false, narrow_crcs = false;
+	struct bpos data_pos = bkey_start_pos(k.k);
+	int pick_ret;
+
+	if (bkey_extent_is_inline_data(k.k)) {
+		unsigned bytes = min_t(unsigned, iter.bi_size,
+				       bkey_inline_data_bytes(k.k));
+
+		swap(iter.bi_size, bytes);
+		memcpy_to_bio(&orig->bio, iter, bkey_inline_data_p(k));
+		swap(iter.bi_size, bytes);
+		bio_advance_iter(&orig->bio, &iter, bytes);
+		zero_fill_bio_iter(&orig->bio, iter);
+		goto out_read_done;
+	}
+retry_pick:
+	pick_ret = bch2_bkey_pick_read_device(c, k, failed, &pick);
+
+	/* hole or reservation - just zero fill: */
+	if (!pick_ret)
+		goto hole;
+
+	if (pick_ret < 0) {
+		bch_err_inum_offset_ratelimited(c,
+				read_pos.inode, read_pos.offset << 9,
+				"no device to read from");
+		goto err;
+	}
+
+	ca = bch_dev_bkey_exists(c, pick.ptr.dev);
+
+	/*
+	 * Stale dirty pointers are treated as IO errors, but @failed isn't
+	 * allocated unless we're in the retry path - so if we're not in the
+	 * retry path, don't check here, it'll be caught in bch2_read_endio()
+	 * and we'll end up in the retry path:
+	 */
+	if ((flags & BCH_READ_IN_RETRY) &&
+	    !pick.ptr.cached &&
+	    unlikely(ptr_stale(ca, &pick.ptr))) {
+		read_from_stale_dirty_pointer(trans, k, pick.ptr);
+		bch2_mark_io_failure(failed, &pick);
+		goto retry_pick;
+	}
+
+	/*
+	 * Unlock the iterator while the btree node's lock is still in
+	 * cache, before doing the IO:
+	 */
+	bch2_trans_unlock(trans);
+
+	if (flags & BCH_READ_NODECODE) {
+		/*
+		 * can happen if we retry, and the extent we were going to read
+		 * has been merged in the meantime:
+		 */
+		if (pick.crc.compressed_size > orig->bio.bi_vcnt * PAGE_SECTORS)
+			goto hole;
+
+		iter.bi_size	= pick.crc.compressed_size << 9;
+		goto get_bio;
+	}
+
+	if (!(flags & BCH_READ_LAST_FRAGMENT) ||
+	    bio_flagged(&orig->bio, BIO_CHAIN))
+		flags |= BCH_READ_MUST_CLONE;
+
+	narrow_crcs = !(flags & BCH_READ_IN_RETRY) &&
+		bch2_can_narrow_extent_crcs(k, pick.crc);
+
+	if (narrow_crcs && (flags & BCH_READ_USER_MAPPED))
+		flags |= BCH_READ_MUST_BOUNCE;
+
+	EBUG_ON(offset_into_extent + bvec_iter_sectors(iter) > k.k->size);
+
+	if (crc_is_compressed(pick.crc) ||
+	    (pick.crc.csum_type != BCH_CSUM_none &&
+	     (bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
+	      (bch2_csum_type_is_encryption(pick.crc.csum_type) &&
+	       (flags & BCH_READ_USER_MAPPED)) ||
+	      (flags & BCH_READ_MUST_BOUNCE)))) {
+		read_full = true;
+		bounce = true;
+	}
+
+	if (orig->opts.promote_target)
+		promote = promote_alloc(trans, iter, k, &pick, orig->opts, flags,
+					&rbio, &bounce, &read_full);
+
+	if (!read_full) {
+		EBUG_ON(crc_is_compressed(pick.crc));
+		EBUG_ON(pick.crc.csum_type &&
+			(bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
+			 bvec_iter_sectors(iter) != pick.crc.live_size ||
+			 pick.crc.offset ||
+			 offset_into_extent));
+
+		data_pos.offset += offset_into_extent;
+		pick.ptr.offset += pick.crc.offset +
+			offset_into_extent;
+		offset_into_extent		= 0;
+		pick.crc.compressed_size	= bvec_iter_sectors(iter);
+		pick.crc.uncompressed_size	= bvec_iter_sectors(iter);
+		pick.crc.offset			= 0;
+		pick.crc.live_size		= bvec_iter_sectors(iter);
+		offset_into_extent		= 0;
+	}
+get_bio:
+	if (rbio) {
+		/*
+		 * promote already allocated bounce rbio:
+		 * promote needs to allocate a bio big enough for uncompressing
+		 * data in the write path, but we're not going to use it all
+		 * here:
+		 */
+		EBUG_ON(rbio->bio.bi_iter.bi_size <
+		       pick.crc.compressed_size << 9);
+		rbio->bio.bi_iter.bi_size =
+			pick.crc.compressed_size << 9;
+	} else if (bounce) {
+		unsigned sectors = pick.crc.compressed_size;
+
+		rbio = rbio_init(bio_alloc_bioset(NULL,
+						  DIV_ROUND_UP(sectors, PAGE_SECTORS),
+						  0,
+						  GFP_NOFS,
+						  &c->bio_read_split),
+				 orig->opts);
+
+		bch2_bio_alloc_pages_pool(c, &rbio->bio, sectors << 9);
+		rbio->bounce	= true;
+		rbio->split	= true;
+	} else if (flags & BCH_READ_MUST_CLONE) {
+		/*
+		 * Have to clone if there were any splits, due to error
+		 * reporting issues (if a split errored, and retrying didn't
+		 * work, when it reports the error to its parent (us) we don't
+		 * know if the error was from our bio, and we should retry, or
+		 * from the whole bio, in which case we don't want to retry and
+		 * lose the error)
+		 */
+		rbio = rbio_init(bio_alloc_clone(NULL, &orig->bio, GFP_NOFS,
+						 &c->bio_read_split),
+				 orig->opts);
+		rbio->bio.bi_iter = iter;
+		rbio->split	= true;
+	} else {
+		rbio = orig;
+		rbio->bio.bi_iter = iter;
+		EBUG_ON(bio_flagged(&rbio->bio, BIO_CHAIN));
+	}
+
+	EBUG_ON(bio_sectors(&rbio->bio) != pick.crc.compressed_size);
+
+	rbio->c			= c;
+	rbio->submit_time	= local_clock();
+	if (rbio->split)
+		rbio->parent	= orig;
+	else
+		rbio->end_io	= orig->bio.bi_end_io;
+	rbio->bvec_iter		= iter;
+	rbio->offset_into_extent= offset_into_extent;
+	rbio->flags		= flags;
+	rbio->have_ioref	= pick_ret > 0 && bch2_dev_get_ioref(ca, READ);
+	rbio->narrow_crcs	= narrow_crcs;
+	rbio->hole		= 0;
+	rbio->retry		= 0;
+	rbio->context		= 0;
+	/* XXX: only initialize this if needed */
+	rbio->devs_have		= bch2_bkey_devs(k);
+	rbio->pick		= pick;
+	rbio->subvol		= orig->subvol;
+	rbio->read_pos		= read_pos;
+	rbio->data_btree	= data_btree;
+	rbio->data_pos		= data_pos;
+	rbio->version		= k.k->version;
+	rbio->promote		= promote;
+	INIT_WORK(&rbio->work, NULL);
+
+	rbio->bio.bi_opf	= orig->bio.bi_opf;
+	rbio->bio.bi_iter.bi_sector = pick.ptr.offset;
+	rbio->bio.bi_end_io	= bch2_read_endio;
+
+	if (rbio->bounce)
+		trace_and_count(c, read_bounce, &rbio->bio);
+
+	this_cpu_add(c->counters[BCH_COUNTER_io_read], bio_sectors(&rbio->bio));
+	bch2_increment_clock(c, bio_sectors(&rbio->bio), READ);
+
+	/*
+	 * If it's being moved internally, we don't want to flag it as a cache
+	 * hit:
+	 */
+	if (pick.ptr.cached && !(flags & BCH_READ_NODECODE))
+		bch2_bucket_io_time_reset(trans, pick.ptr.dev,
+			PTR_BUCKET_NR(ca, &pick.ptr), READ);
+
+	if (!(flags & (BCH_READ_IN_RETRY|BCH_READ_LAST_FRAGMENT))) {
+		bio_inc_remaining(&orig->bio);
+		trace_and_count(c, read_split, &orig->bio);
+	}
+
+	if (!rbio->pick.idx) {
+		if (!rbio->have_ioref) {
+			bch_err_inum_offset_ratelimited(c,
+					read_pos.inode,
+					read_pos.offset << 9,
+					"no device to read from");
+			bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
+			goto out;
+		}
+
+		this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_user],
+			     bio_sectors(&rbio->bio));
+		bio_set_dev(&rbio->bio, ca->disk_sb.bdev);
+
+		if (unlikely(c->opts.no_data_io)) {
+			if (likely(!(flags & BCH_READ_IN_RETRY)))
+				bio_endio(&rbio->bio);
+		} else {
+			if (likely(!(flags & BCH_READ_IN_RETRY)))
+				submit_bio(&rbio->bio);
+			else
+				submit_bio_wait(&rbio->bio);
+		}
+
+		/*
+		 * We just submitted IO which may block, we expect relock fail
+		 * events and shouldn't count them:
+		 */
+		trans->notrace_relock_fail = true;
+	} else {
+		/* Attempting reconstruct read: */
+		if (bch2_ec_read_extent(c, rbio)) {
+			bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
+			goto out;
+		}
+
+		if (likely(!(flags & BCH_READ_IN_RETRY)))
+			bio_endio(&rbio->bio);
+	}
+out:
+	if (likely(!(flags & BCH_READ_IN_RETRY))) {
+		return 0;
+	} else {
+		int ret;
+
+		rbio->context = RBIO_CONTEXT_UNBOUND;
+		bch2_read_endio(&rbio->bio);
+
+		ret = rbio->retry;
+		rbio = bch2_rbio_free(rbio);
+
+		if (ret == READ_RETRY_AVOID) {
+			bch2_mark_io_failure(failed, &pick);
+			ret = READ_RETRY;
+		}
+
+		if (!ret)
+			goto out_read_done;
+
+		return ret;
+	}
+
+err:
+	if (flags & BCH_READ_IN_RETRY)
+		return READ_ERR;
+
+	orig->bio.bi_status = BLK_STS_IOERR;
+	goto out_read_done;
+
+hole:
+	/*
+	 * won't normally happen in the BCH_READ_NODECODE
+	 * (bch2_move_extent()) path, but if we retry and the extent we wanted
+	 * to read no longer exists we have to signal that:
+	 */
+	if (flags & BCH_READ_NODECODE)
+		orig->hole = true;
+
+	zero_fill_bio_iter(&orig->bio, iter);
+out_read_done:
+	if (flags & BCH_READ_LAST_FRAGMENT)
+		bch2_rbio_done(orig);
+	return 0;
+}
+
+void __bch2_read(struct bch_fs *c, struct bch_read_bio *rbio,
+		 struct bvec_iter bvec_iter, subvol_inum inum,
+		 struct bch_io_failures *failed, unsigned flags)
+{
+	struct btree_trans trans;
+	struct btree_iter iter;
+	struct bkey_buf sk;
+	struct bkey_s_c k;
+	u32 snapshot;
+	int ret;
+
+	BUG_ON(flags & BCH_READ_NODECODE);
+
+	bch2_bkey_buf_init(&sk);
+	bch2_trans_init(&trans, c, 0, 0);
+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, bvec_iter.bi_sector, snapshot),
+			     BTREE_ITER_SLOTS);
+	while (1) {
+		unsigned bytes, sectors, offset_into_extent;
+		enum btree_id data_btree = BTREE_ID_extents;
+
+		/*
+		 * 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;
+
+		bch2_btree_iter_set_pos(&iter,
+				POS(inum.inum, bvec_iter.bi_sector));
+
+		k = bch2_btree_iter_peek_slot(&iter);
+		ret = bkey_err(k);
+		if (ret)
+			break;
+
+		offset_into_extent = iter.pos.offset -
+			bkey_start_offset(k.k);
+		sectors = k.k->size - offset_into_extent;
+
+		bch2_bkey_buf_reassemble(&sk, c, k);
+
+		ret = bch2_read_indirect_extent(&trans, &data_btree,
+					&offset_into_extent, &sk);
+		if (ret)
+			break;
+
+		k = bkey_i_to_s_c(sk.k);
+
+		/*
+		 * With indirect extents, the amount of data to read is the min
+		 * of the original extent and the indirect extent:
+		 */
+		sectors = min(sectors, k.k->size - offset_into_extent);
+
+		bytes = min(sectors, bvec_iter_sectors(bvec_iter)) << 9;
+		swap(bvec_iter.bi_size, bytes);
+
+		if (bvec_iter.bi_size == bytes)
+			flags |= BCH_READ_LAST_FRAGMENT;
+
+		ret = __bch2_read_extent(&trans, rbio, bvec_iter, iter.pos,
+					 data_btree, k,
+					 offset_into_extent, failed, flags);
+		if (ret)
+			break;
+
+		if (flags & BCH_READ_LAST_FRAGMENT)
+			break;
+
+		swap(bvec_iter.bi_size, bytes);
+		bio_advance_iter(&rbio->bio, &bvec_iter, 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) ||
+	    ret == READ_RETRY ||
+	    ret == READ_RETRY_AVOID)
+		goto retry;
+
+	bch2_trans_exit(&trans);
+	bch2_bkey_buf_exit(&sk, c);
+
+	if (ret) {
+		bch_err_inum_offset_ratelimited(c, inum.inum,
+						bvec_iter.bi_sector << 9,
+						"read error %i from btree lookup", ret);
+		rbio->bio.bi_status = BLK_STS_IOERR;
+		bch2_rbio_done(rbio);
+	}
+}
+
+void bch2_fs_io_read_exit(struct bch_fs *c)
+{
+	if (c->promote_table.tbl)
+		rhashtable_destroy(&c->promote_table);
+	bioset_exit(&c->bio_read_split);
+	bioset_exit(&c->bio_read);
+}
+
+int bch2_fs_io_read_init(struct bch_fs *c)
+{
+	if (bioset_init(&c->bio_read, 1, offsetof(struct bch_read_bio, bio),
+			BIOSET_NEED_BVECS))
+		return -BCH_ERR_ENOMEM_bio_read_init;
+
+	if (bioset_init(&c->bio_read_split, 1, offsetof(struct bch_read_bio, bio),
+			BIOSET_NEED_BVECS))
+		return -BCH_ERR_ENOMEM_bio_read_split_init;
+
+	if (rhashtable_init(&c->promote_table, &bch_promote_params))
+		return -BCH_ERR_ENOMEM_promote_table_init;
+
+	return 0;
+}
diff --git a/fs/bcachefs/io_read.h b/fs/bcachefs/io_read.h
new file mode 100644
index 000000000000..d9c18bb7d403
--- /dev/null
+++ b/fs/bcachefs/io_read.h
@@ -0,0 +1,158 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _BCACHEFS_IO_READ_H
+#define _BCACHEFS_IO_READ_H
+
+#include "bkey_buf.h"
+
+struct bch_read_bio {
+	struct bch_fs		*c;
+	u64			start_time;
+	u64			submit_time;
+
+	/*
+	 * Reads will often have to be split, and if the extent being read from
+	 * was checksummed or compressed we'll also have to allocate bounce
+	 * buffers and copy the data back into the original bio.
+	 *
+	 * If we didn't have to split, we have to save and restore the original
+	 * bi_end_io - @split below indicates which:
+	 */
+	union {
+	struct bch_read_bio	*parent;
+	bio_end_io_t		*end_io;
+	};
+
+	/*
+	 * Saved copy of bio->bi_iter, from submission time - allows us to
+	 * resubmit on IO error, and also to copy data back to the original bio
+	 * when we're bouncing:
+	 */
+	struct bvec_iter	bvec_iter;
+
+	unsigned		offset_into_extent;
+
+	u16			flags;
+	union {
+	struct {
+	u16			bounce:1,
+				split:1,
+				kmalloc:1,
+				have_ioref:1,
+				narrow_crcs:1,
+				hole:1,
+				retry:2,
+				context:2;
+	};
+	u16			_state;
+	};
+
+	struct bch_devs_list	devs_have;
+
+	struct extent_ptr_decoded pick;
+
+	/*
+	 * pos we read from - different from data_pos for indirect extents:
+	 */
+	u32			subvol;
+	struct bpos		read_pos;
+
+	/*
+	 * start pos of data we read (may not be pos of data we want) - for
+	 * promote, narrow extents paths:
+	 */
+	enum btree_id		data_btree;
+	struct bpos		data_pos;
+	struct bversion		version;
+
+	struct promote_op	*promote;
+
+	struct bch_io_opts	opts;
+
+	struct work_struct	work;
+
+	struct bio		bio;
+};
+
+#define to_rbio(_bio)		container_of((_bio), struct bch_read_bio, bio)
+
+struct bch_devs_mask;
+struct cache_promote_op;
+struct extent_ptr_decoded;
+
+int __bch2_read_indirect_extent(struct btree_trans *, unsigned *,
+				struct bkey_buf *);
+
+static inline int bch2_read_indirect_extent(struct btree_trans *trans,
+					    enum btree_id *data_btree,
+					    unsigned *offset_into_extent,
+					    struct bkey_buf *k)
+{
+	if (k->k->k.type != KEY_TYPE_reflink_p)
+		return 0;
+
+	*data_btree = BTREE_ID_reflink;
+	return __bch2_read_indirect_extent(trans, offset_into_extent, k);
+}
+
+enum bch_read_flags {
+	BCH_READ_RETRY_IF_STALE		= 1 << 0,
+	BCH_READ_MAY_PROMOTE		= 1 << 1,
+	BCH_READ_USER_MAPPED		= 1 << 2,
+	BCH_READ_NODECODE		= 1 << 3,
+	BCH_READ_LAST_FRAGMENT		= 1 << 4,
+
+	/* internal: */
+	BCH_READ_MUST_BOUNCE		= 1 << 5,
+	BCH_READ_MUST_CLONE		= 1 << 6,
+	BCH_READ_IN_RETRY		= 1 << 7,
+};
+
+int __bch2_read_extent(struct btree_trans *, struct bch_read_bio *,
+		       struct bvec_iter, struct bpos, enum btree_id,
+		       struct bkey_s_c, unsigned,
+		       struct bch_io_failures *, unsigned);
+
+static inline void bch2_read_extent(struct btree_trans *trans,
+			struct bch_read_bio *rbio, struct bpos read_pos,
+			enum btree_id data_btree, struct bkey_s_c k,
+			unsigned offset_into_extent, unsigned flags)
+{
+	__bch2_read_extent(trans, rbio, rbio->bio.bi_iter, read_pos,
+			   data_btree, k, offset_into_extent, NULL, flags);
+}
+
+void __bch2_read(struct bch_fs *, struct bch_read_bio *, struct bvec_iter,
+		 subvol_inum, struct bch_io_failures *, unsigned flags);
+
+static inline void bch2_read(struct bch_fs *c, struct bch_read_bio *rbio,
+			     subvol_inum inum)
+{
+	struct bch_io_failures failed = { .nr = 0 };
+
+	BUG_ON(rbio->_state);
+
+	rbio->c = c;
+	rbio->start_time = local_clock();
+	rbio->subvol = inum.subvol;
+
+	__bch2_read(c, rbio, rbio->bio.bi_iter, inum, &failed,
+		    BCH_READ_RETRY_IF_STALE|
+		    BCH_READ_MAY_PROMOTE|
+		    BCH_READ_USER_MAPPED);
+}
+
+static inline struct bch_read_bio *rbio_init(struct bio *bio,
+					     struct bch_io_opts opts)
+{
+	struct bch_read_bio *rbio = to_rbio(bio);
+
+	rbio->_state	= 0;
+	rbio->promote	= NULL;
+	rbio->opts	= opts;
+	return rbio;
+}
+
+void bch2_fs_io_read_exit(struct bch_fs *);
+int bch2_fs_io_read_init(struct bch_fs *);
+
+#endif /* _BCACHEFS_IO_READ_H */
diff --git a/fs/bcachefs/io.c b/fs/bcachefs/io_write.c
index 3c614c864b6e..7f29fd2f05b1 100644
--- a/fs/bcachefs/io.c
+++ b/fs/bcachefs/io_write.c
@@ -1,29 +1,24 @@
 // SPDX-License-Identifier: GPL-2.0
 /*
- * Some low level IO code, and hacks for various block layer limitations
- *
  * Copyright 2010, 2011 Kent Overstreet <kent.overstreet@gmail.com>
  * Copyright 2012 Google, Inc.
  */
 
 #include "bcachefs.h"
-#include "alloc_background.h"
 #include "alloc_foreground.h"
 #include "bkey_buf.h"
 #include "bset.h"
 #include "btree_update.h"
 #include "buckets.h"
 #include "checksum.h"
-#include "compress.h"
 #include "clock.h"
-#include "data_update.h"
+#include "compress.h"
 #include "debug.h"
-#include "disk_groups.h"
 #include "ec.h"
 #include "error.h"
 #include "extent_update.h"
 #include "inode.h"
-#include "io.h"
+#include "io_write.h"
 #include "journal.h"
 #include "keylist.h"
 #include "move.h"
@@ -39,48 +34,8 @@
 #include <linux/random.h>
 #include <linux/sched/mm.h>
 
-const char *bch2_blk_status_to_str(blk_status_t status)
-{
-	if (status == BLK_STS_REMOVED)
-		return "device removed";
-	return blk_status_to_str(status);
-}
-
 #ifndef CONFIG_BCACHEFS_NO_LATENCY_ACCT
 
-static bool bch2_target_congested(struct bch_fs *c, u16 target)
-{
-	const struct bch_devs_mask *devs;
-	unsigned d, nr = 0, total = 0;
-	u64 now = local_clock(), last;
-	s64 congested;
-	struct bch_dev *ca;
-
-	if (!target)
-		return false;
-
-	rcu_read_lock();
-	devs = bch2_target_to_mask(c, target) ?:
-		&c->rw_devs[BCH_DATA_user];
-
-	for_each_set_bit(d, devs->d, BCH_SB_MEMBERS_MAX) {
-		ca = rcu_dereference(c->devs[d]);
-		if (!ca)
-			continue;
-
-		congested = atomic_read(&ca->congested);
-		last = READ_ONCE(ca->congested_last);
-		if (time_after64(now, last))
-			congested -= (now - last) >> 12;
-
-		total += max(congested, 0LL);
-		nr++;
-	}
-	rcu_read_unlock();
-
-	return bch2_rand_range(nr * CONGESTED_MAX) < total;
-}
-
 static inline void bch2_congested_acct(struct bch_dev *ca, u64 io_latency,
 				       u64 now, int rw)
 {
@@ -136,13 +91,6 @@ void bch2_latency_acct(struct bch_dev *ca, u64 submit_time, int rw)
 	__bch2_time_stats_update(&ca->io_latency[rw], submit_time, now);
 }
 
-#else
-
-static bool bch2_target_congested(struct bch_fs *c, u16 target)
-{
-	return false;
-}
-
 #endif
 
 /* Allocate, free from mempool: */
@@ -368,206 +316,6 @@ int bch2_extent_update(struct btree_trans *trans,
 	return 0;
 }
 
-/* Overwrites whatever was present with zeroes: */
-int bch2_extent_fallocate(struct btree_trans *trans,
-			  subvol_inum inum,
-			  struct btree_iter *iter,
-			  unsigned sectors,
-			  struct bch_io_opts opts,
-			  s64 *i_sectors_delta,
-			  struct write_point_specifier write_point)
-{
-	struct bch_fs *c = trans->c;
-	struct disk_reservation disk_res = { 0 };
-	struct closure cl;
-	struct open_buckets open_buckets = { 0 };
-	struct bkey_s_c k;
-	struct bkey_buf old, new;
-	unsigned sectors_allocated = 0;
-	bool have_reservation = false;
-	bool unwritten = opts.nocow &&
-	    c->sb.version >= bcachefs_metadata_version_unwritten_extents;
-	int ret;
-
-	bch2_bkey_buf_init(&old);
-	bch2_bkey_buf_init(&new);
-	closure_init_stack(&cl);
-
-	k = bch2_btree_iter_peek_slot(iter);
-	ret = bkey_err(k);
-	if (ret)
-		return ret;
-
-	sectors = min_t(u64, sectors, k.k->p.offset - iter->pos.offset);
-
-	if (!have_reservation) {
-		unsigned new_replicas =
-			max(0, (int) opts.data_replicas -
-			    (int) bch2_bkey_nr_ptrs_fully_allocated(k));
-		/*
-		 * Get a disk reservation before (in the nocow case) calling
-		 * into the allocator:
-		 */
-		ret = bch2_disk_reservation_get(c, &disk_res, sectors, new_replicas, 0);
-		if (unlikely(ret))
-			goto err;
-
-		bch2_bkey_buf_reassemble(&old, c, k);
-	}
-
-	if (have_reservation) {
-		if (!bch2_extents_match(k, bkey_i_to_s_c(old.k)))
-			goto err;
-
-		bch2_key_resize(&new.k->k, sectors);
-	} else if (!unwritten) {
-		struct bkey_i_reservation *reservation;
-
-		bch2_bkey_buf_realloc(&new, c, sizeof(*reservation) / sizeof(u64));
-		reservation = bkey_reservation_init(new.k);
-		reservation->k.p = iter->pos;
-		bch2_key_resize(&reservation->k, sectors);
-		reservation->v.nr_replicas = opts.data_replicas;
-	} else {
-		struct bkey_i_extent *e;
-		struct bch_devs_list devs_have;
-		struct write_point *wp;
-		struct bch_extent_ptr *ptr;
-
-		devs_have.nr = 0;
-
-		bch2_bkey_buf_realloc(&new, c, BKEY_EXTENT_U64s_MAX);
-
-		e = bkey_extent_init(new.k);
-		e->k.p = iter->pos;
-
-		ret = bch2_alloc_sectors_start_trans(trans,
-				opts.foreground_target,
-				false,
-				write_point,
-				&devs_have,
-				opts.data_replicas,
-				opts.data_replicas,
-				BCH_WATERMARK_normal, 0, &cl, &wp);
-		if (bch2_err_matches(ret, BCH_ERR_operation_blocked))
-			ret = -BCH_ERR_transaction_restart_nested;
-		if (ret)
-			goto err;
-
-		sectors = min(sectors, wp->sectors_free);
-		sectors_allocated = sectors;
-
-		bch2_key_resize(&e->k, sectors);
-
-		bch2_open_bucket_get(c, wp, &open_buckets);
-		bch2_alloc_sectors_append_ptrs(c, wp, &e->k_i, sectors, false);
-		bch2_alloc_sectors_done(c, wp);
-
-		extent_for_each_ptr(extent_i_to_s(e), ptr)
-			ptr->unwritten = true;
-	}
-
-	have_reservation = true;
-
-	ret = bch2_extent_update(trans, inum, iter, new.k, &disk_res,
-				 0, i_sectors_delta, true);
-err:
-	if (!ret && sectors_allocated)
-		bch2_increment_clock(c, sectors_allocated, WRITE);
-
-	bch2_open_buckets_put(c, &open_buckets);
-	bch2_disk_reservation_put(c, &disk_res);
-	bch2_bkey_buf_exit(&new, c);
-	bch2_bkey_buf_exit(&old, c);
-
-	if (closure_nr_remaining(&cl) != 1) {
-		bch2_trans_unlock(trans);
-		closure_sync(&cl);
-	}
-
-	return ret;
-}
-
-/*
- * Returns -BCH_ERR_transacton_restart if we had to drop locks:
- */
-int bch2_fpunch_at(struct btree_trans *trans, struct btree_iter *iter,
-		   subvol_inum inum, u64 end,
-		   s64 *i_sectors_delta)
-{
-	struct bch_fs *c	= trans->c;
-	unsigned max_sectors	= KEY_SIZE_MAX & (~0 << c->block_bits);
-	struct bpos end_pos = POS(inum.inum, end);
-	struct bkey_s_c k;
-	int ret = 0, ret2 = 0;
-	u32 snapshot;
-
-	while (!ret ||
-	       bch2_err_matches(ret, BCH_ERR_transaction_restart)) {
-		struct disk_reservation disk_res =
-			bch2_disk_reservation_init(c, 0);
-		struct bkey_i delete;
-
-		if (ret)
-			ret2 = ret;
-
-		bch2_trans_begin(trans);
-
-		ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
-		if (ret)
-			continue;
-
-		bch2_btree_iter_set_snapshot(iter, snapshot);
-
-		/*
-		 * peek_upto() doesn't have ideal semantics for extents:
-		 */
-		k = bch2_btree_iter_peek_upto(iter, end_pos);
-		if (!k.k)
-			break;
-
-		ret = bkey_err(k);
-		if (ret)
-			continue;
-
-		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_pos, &delete);
-
-		ret = bch2_extent_update(trans, inum, iter, &delete,
-				&disk_res, 0, i_sectors_delta, false);
-		bch2_disk_reservation_put(c, &disk_res);
-	}
-
-	return ret ?: ret2;
-}
-
-int bch2_fpunch(struct bch_fs *c, subvol_inum inum, u64 start, u64 end,
-		s64 *i_sectors_delta)
-{
-	struct btree_trans trans;
-	struct btree_iter iter;
-	int ret;
-
-	bch2_trans_init(&trans, c, BTREE_ITER_MAX, 1024);
-	bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents,
-			     POS(inum.inum, start),
-			     BTREE_ITER_INTENT);
-
-	ret = bch2_fpunch_at(&trans, &iter, inum, end, i_sectors_delta);
-
-	bch2_trans_iter_exit(&trans, &iter);
-	bch2_trans_exit(&trans);
-
-	if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
-		ret = 0;
-
-	return ret;
-}
-
 static int bch2_write_index_default(struct bch_write_op *op)
 {
 	struct bch_fs *c = op->c;
@@ -1899,1140 +1647,14 @@ void bch2_write_op_to_text(struct printbuf *out, struct bch_write_op *op)
 	printbuf_indent_sub(out, 2);
 }
 
-/* Cache promotion on read */
-
-struct promote_op {
-	struct rcu_head		rcu;
-	u64			start_time;
-
-	struct rhash_head	hash;
-	struct bpos		pos;
-
-	struct data_update	write;
-	struct bio_vec		bi_inline_vecs[0]; /* must be last */
-};
-
-static const struct rhashtable_params bch_promote_params = {
-	.head_offset	= offsetof(struct promote_op, hash),
-	.key_offset	= offsetof(struct promote_op, pos),
-	.key_len	= sizeof(struct bpos),
-};
-
-static inline bool should_promote(struct bch_fs *c, struct bkey_s_c k,
-				  struct bpos pos,
-				  struct bch_io_opts opts,
-				  unsigned flags)
-{
-	if (!(flags & BCH_READ_MAY_PROMOTE))
-		return false;
-
-	if (!opts.promote_target)
-		return false;
-
-	if (bch2_bkey_has_target(c, k, opts.promote_target))
-		return false;
-
-	if (bkey_extent_is_unwritten(k))
-		return false;
-
-	if (bch2_target_congested(c, opts.promote_target)) {
-		/* XXX trace this */
-		return false;
-	}
-
-	if (rhashtable_lookup_fast(&c->promote_table, &pos,
-				   bch_promote_params))
-		return false;
-
-	return true;
-}
-
-static void promote_free(struct bch_fs *c, struct promote_op *op)
-{
-	int ret;
-
-	bch2_data_update_exit(&op->write);
-
-	ret = rhashtable_remove_fast(&c->promote_table, &op->hash,
-				     bch_promote_params);
-	BUG_ON(ret);
-	bch2_write_ref_put(c, BCH_WRITE_REF_promote);
-	kfree_rcu(op, rcu);
-}
-
-static void promote_done(struct bch_write_op *wop)
-{
-	struct promote_op *op =
-		container_of(wop, struct promote_op, write.op);
-	struct bch_fs *c = op->write.op.c;
-
-	bch2_time_stats_update(&c->times[BCH_TIME_data_promote],
-			       op->start_time);
-	promote_free(c, op);
-}
-
-static void promote_start(struct promote_op *op, struct bch_read_bio *rbio)
-{
-	struct bio *bio = &op->write.op.wbio.bio;
-
-	trace_and_count(op->write.op.c, read_promote, &rbio->bio);
-
-	/* we now own pages: */
-	BUG_ON(!rbio->bounce);
-	BUG_ON(rbio->bio.bi_vcnt > bio->bi_max_vecs);
-
-	memcpy(bio->bi_io_vec, rbio->bio.bi_io_vec,
-	       sizeof(struct bio_vec) * rbio->bio.bi_vcnt);
-	swap(bio->bi_vcnt, rbio->bio.bi_vcnt);
-
-	bch2_data_update_read_done(&op->write, rbio->pick.crc);
-}
-
-static struct promote_op *__promote_alloc(struct btree_trans *trans,
-					  enum btree_id btree_id,
-					  struct bkey_s_c k,
-					  struct bpos pos,
-					  struct extent_ptr_decoded *pick,
-					  struct bch_io_opts opts,
-					  unsigned sectors,
-					  struct bch_read_bio **rbio)
-{
-	struct bch_fs *c = trans->c;
-	struct promote_op *op = NULL;
-	struct bio *bio;
-	unsigned pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
-	int ret;
-
-	if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_promote))
-		return NULL;
-
-	op = kzalloc(sizeof(*op) + sizeof(struct bio_vec) * pages, GFP_NOFS);
-	if (!op)
-		goto err;
-
-	op->start_time = local_clock();
-	op->pos = pos;
-
-	/*
-	 * We don't use the mempool here because extents that aren't
-	 * checksummed or compressed can be too big for the mempool:
-	 */
-	*rbio = kzalloc(sizeof(struct bch_read_bio) +
-			sizeof(struct bio_vec) * pages,
-			GFP_NOFS);
-	if (!*rbio)
-		goto err;
-
-	rbio_init(&(*rbio)->bio, opts);
-	bio_init(&(*rbio)->bio, NULL, (*rbio)->bio.bi_inline_vecs, pages, 0);
-
-	if (bch2_bio_alloc_pages(&(*rbio)->bio, sectors << 9,
-				 GFP_NOFS))
-		goto err;
-
-	(*rbio)->bounce		= true;
-	(*rbio)->split		= true;
-	(*rbio)->kmalloc	= true;
-
-	if (rhashtable_lookup_insert_fast(&c->promote_table, &op->hash,
-					  bch_promote_params))
-		goto err;
-
-	bio = &op->write.op.wbio.bio;
-	bio_init(bio, NULL, bio->bi_inline_vecs, pages, 0);
-
-	ret = bch2_data_update_init(trans, NULL, &op->write,
-			writepoint_hashed((unsigned long) current),
-			opts,
-			(struct data_update_opts) {
-				.target		= opts.promote_target,
-				.extra_replicas	= 1,
-				.write_flags	= BCH_WRITE_ALLOC_NOWAIT|BCH_WRITE_CACHED,
-			},
-			btree_id, k);
-	/*
-	 * possible errors: -BCH_ERR_nocow_lock_blocked,
-	 * -BCH_ERR_ENOSPC_disk_reservation:
-	 */
-	if (ret) {
-		ret = rhashtable_remove_fast(&c->promote_table, &op->hash,
-					bch_promote_params);
-		BUG_ON(ret);
-		goto err;
-	}
-
-	op->write.op.end_io = promote_done;
-
-	return op;
-err:
-	if (*rbio)
-		bio_free_pages(&(*rbio)->bio);
-	kfree(*rbio);
-	*rbio = NULL;
-	kfree(op);
-	bch2_write_ref_put(c, BCH_WRITE_REF_promote);
-	return NULL;
-}
-
-noinline
-static struct promote_op *promote_alloc(struct btree_trans *trans,
-					struct bvec_iter iter,
-					struct bkey_s_c k,
-					struct extent_ptr_decoded *pick,
-					struct bch_io_opts opts,
-					unsigned flags,
-					struct bch_read_bio **rbio,
-					bool *bounce,
-					bool *read_full)
-{
-	struct bch_fs *c = trans->c;
-	bool promote_full = *read_full || READ_ONCE(c->promote_whole_extents);
-	/* data might have to be decompressed in the write path: */
-	unsigned sectors = promote_full
-		? max(pick->crc.compressed_size, pick->crc.live_size)
-		: bvec_iter_sectors(iter);
-	struct bpos pos = promote_full
-		? bkey_start_pos(k.k)
-		: POS(k.k->p.inode, iter.bi_sector);
-	struct promote_op *promote;
-
-	if (!should_promote(c, k, pos, opts, flags))
-		return NULL;
-
-	promote = __promote_alloc(trans,
-				  k.k->type == KEY_TYPE_reflink_v
-				  ? BTREE_ID_reflink
-				  : BTREE_ID_extents,
-				  k, pos, pick, opts, sectors, rbio);
-	if (!promote)
-		return NULL;
-
-	*bounce		= true;
-	*read_full	= promote_full;
-	return promote;
-}
-
-/* Read */
-
-#define READ_RETRY_AVOID	1
-#define READ_RETRY		2
-#define READ_ERR		3
-
-enum rbio_context {
-	RBIO_CONTEXT_NULL,
-	RBIO_CONTEXT_HIGHPRI,
-	RBIO_CONTEXT_UNBOUND,
-};
-
-static inline struct bch_read_bio *
-bch2_rbio_parent(struct bch_read_bio *rbio)
-{
-	return rbio->split ? rbio->parent : rbio;
-}
-
-__always_inline
-static void bch2_rbio_punt(struct bch_read_bio *rbio, work_func_t fn,
-			   enum rbio_context context,
-			   struct workqueue_struct *wq)
-{
-	if (context <= rbio->context) {
-		fn(&rbio->work);
-	} else {
-		rbio->work.func		= fn;
-		rbio->context		= context;
-		queue_work(wq, &rbio->work);
-	}
-}
-
-static inline struct bch_read_bio *bch2_rbio_free(struct bch_read_bio *rbio)
-{
-	BUG_ON(rbio->bounce && !rbio->split);
-
-	if (rbio->promote)
-		promote_free(rbio->c, rbio->promote);
-	rbio->promote = NULL;
-
-	if (rbio->bounce)
-		bch2_bio_free_pages_pool(rbio->c, &rbio->bio);
-
-	if (rbio->split) {
-		struct bch_read_bio *parent = rbio->parent;
-
-		if (rbio->kmalloc)
-			kfree(rbio);
-		else
-			bio_put(&rbio->bio);
-
-		rbio = parent;
-	}
-
-	return rbio;
-}
-
-/*
- * Only called on a top level bch_read_bio to complete an entire read request,
- * not a split:
- */
-static void bch2_rbio_done(struct bch_read_bio *rbio)
-{
-	if (rbio->start_time)
-		bch2_time_stats_update(&rbio->c->times[BCH_TIME_data_read],
-				       rbio->start_time);
-	bio_endio(&rbio->bio);
-}
-
-static void bch2_read_retry_nodecode(struct bch_fs *c, struct bch_read_bio *rbio,
-				     struct bvec_iter bvec_iter,
-				     struct bch_io_failures *failed,
-				     unsigned flags)
-{
-	struct btree_trans trans;
-	struct btree_iter iter;
-	struct bkey_buf sk;
-	struct bkey_s_c k;
-	int ret;
-
-	flags &= ~BCH_READ_LAST_FRAGMENT;
-	flags |= BCH_READ_MUST_CLONE;
-
-	bch2_bkey_buf_init(&sk);
-	bch2_trans_init(&trans, c, 0, 0);
-
-	bch2_trans_iter_init(&trans, &iter, rbio->data_btree,
-			     rbio->read_pos, BTREE_ITER_SLOTS);
-retry:
-	rbio->bio.bi_status = 0;
-
-	k = bch2_btree_iter_peek_slot(&iter);
-	if (bkey_err(k))
-		goto err;
-
-	bch2_bkey_buf_reassemble(&sk, c, k);
-	k = bkey_i_to_s_c(sk.k);
-	bch2_trans_unlock(&trans);
-
-	if (!bch2_bkey_matches_ptr(c, k,
-				   rbio->pick.ptr,
-				   rbio->data_pos.offset -
-				   rbio->pick.crc.offset)) {
-		/* extent we wanted to read no longer exists: */
-		rbio->hole = true;
-		goto out;
-	}
-
-	ret = __bch2_read_extent(&trans, rbio, bvec_iter,
-				 rbio->read_pos,
-				 rbio->data_btree,
-				 k, 0, failed, flags);
-	if (ret == READ_RETRY)
-		goto retry;
-	if (ret)
-		goto err;
-out:
-	bch2_rbio_done(rbio);
-	bch2_trans_iter_exit(&trans, &iter);
-	bch2_trans_exit(&trans);
-	bch2_bkey_buf_exit(&sk, c);
-	return;
-err:
-	rbio->bio.bi_status = BLK_STS_IOERR;
-	goto out;
-}
-
-static void bch2_rbio_retry(struct work_struct *work)
-{
-	struct bch_read_bio *rbio =
-		container_of(work, struct bch_read_bio, work);
-	struct bch_fs *c	= rbio->c;
-	struct bvec_iter iter	= rbio->bvec_iter;
-	unsigned flags		= rbio->flags;
-	subvol_inum inum = {
-		.subvol = rbio->subvol,
-		.inum	= rbio->read_pos.inode,
-	};
-	struct bch_io_failures failed = { .nr = 0 };
-
-	trace_and_count(c, read_retry, &rbio->bio);
-
-	if (rbio->retry == READ_RETRY_AVOID)
-		bch2_mark_io_failure(&failed, &rbio->pick);
-
-	rbio->bio.bi_status = 0;
-
-	rbio = bch2_rbio_free(rbio);
-
-	flags |= BCH_READ_IN_RETRY;
-	flags &= ~BCH_READ_MAY_PROMOTE;
-
-	if (flags & BCH_READ_NODECODE) {
-		bch2_read_retry_nodecode(c, rbio, iter, &failed, flags);
-	} else {
-		flags &= ~BCH_READ_LAST_FRAGMENT;
-		flags |= BCH_READ_MUST_CLONE;
-
-		__bch2_read(c, rbio, iter, inum, &failed, flags);
-	}
-}
-
-static void bch2_rbio_error(struct bch_read_bio *rbio, int retry,
-			    blk_status_t error)
-{
-	rbio->retry = retry;
-
-	if (rbio->flags & BCH_READ_IN_RETRY)
-		return;
-
-	if (retry == READ_ERR) {
-		rbio = bch2_rbio_free(rbio);
-
-		rbio->bio.bi_status = error;
-		bch2_rbio_done(rbio);
-	} else {
-		bch2_rbio_punt(rbio, bch2_rbio_retry,
-			       RBIO_CONTEXT_UNBOUND, system_unbound_wq);
-	}
-}
-
-static int __bch2_rbio_narrow_crcs(struct btree_trans *trans,
-				   struct bch_read_bio *rbio)
-{
-	struct bch_fs *c = rbio->c;
-	u64 data_offset = rbio->data_pos.offset - rbio->pick.crc.offset;
-	struct bch_extent_crc_unpacked new_crc;
-	struct btree_iter iter;
-	struct bkey_i *new;
-	struct bkey_s_c k;
-	int ret = 0;
-
-	if (crc_is_compressed(rbio->pick.crc))
-		return 0;
-
-	k = bch2_bkey_get_iter(trans, &iter, rbio->data_btree, rbio->data_pos,
-			       BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
-	if ((ret = bkey_err(k)))
-		goto out;
-
-	if (bversion_cmp(k.k->version, rbio->version) ||
-	    !bch2_bkey_matches_ptr(c, k, rbio->pick.ptr, data_offset))
-		goto out;
-
-	/* Extent was merged? */
-	if (bkey_start_offset(k.k) < data_offset ||
-	    k.k->p.offset > data_offset + rbio->pick.crc.uncompressed_size)
-		goto out;
-
-	if (bch2_rechecksum_bio(c, &rbio->bio, rbio->version,
-			rbio->pick.crc, NULL, &new_crc,
-			bkey_start_offset(k.k) - data_offset, k.k->size,
-			rbio->pick.crc.csum_type)) {
-		bch_err(c, "error verifying existing checksum while narrowing checksum (memory corruption?)");
-		ret = 0;
-		goto out;
-	}
-
-	/*
-	 * going to be temporarily appending another checksum entry:
-	 */
-	new = bch2_trans_kmalloc(trans, bkey_bytes(k.k) +
-				 sizeof(struct bch_extent_crc128));
-	if ((ret = PTR_ERR_OR_ZERO(new)))
-		goto out;
-
-	bkey_reassemble(new, k);
-
-	if (!bch2_bkey_narrow_crcs(new, new_crc))
-		goto out;
-
-	ret = bch2_trans_update(trans, &iter, new,
-				BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
-out:
-	bch2_trans_iter_exit(trans, &iter);
-	return ret;
-}
-
-static noinline void bch2_rbio_narrow_crcs(struct bch_read_bio *rbio)
-{
-	bch2_trans_do(rbio->c, NULL, NULL, BTREE_INSERT_NOFAIL,
-		      __bch2_rbio_narrow_crcs(&trans, rbio));
-}
-
-/* Inner part that may run in process context */
-static void __bch2_read_endio(struct work_struct *work)
-{
-	struct bch_read_bio *rbio =
-		container_of(work, struct bch_read_bio, work);
-	struct bch_fs *c	= rbio->c;
-	struct bch_dev *ca	= bch_dev_bkey_exists(c, rbio->pick.ptr.dev);
-	struct bio *src		= &rbio->bio;
-	struct bio *dst		= &bch2_rbio_parent(rbio)->bio;
-	struct bvec_iter dst_iter = rbio->bvec_iter;
-	struct bch_extent_crc_unpacked crc = rbio->pick.crc;
-	struct nonce nonce = extent_nonce(rbio->version, crc);
-	unsigned nofs_flags;
-	struct bch_csum csum;
-	int ret;
-
-	nofs_flags = memalloc_nofs_save();
-
-	/* Reset iterator for checksumming and copying bounced data: */
-	if (rbio->bounce) {
-		src->bi_iter.bi_size		= crc.compressed_size << 9;
-		src->bi_iter.bi_idx		= 0;
-		src->bi_iter.bi_bvec_done	= 0;
-	} else {
-		src->bi_iter			= rbio->bvec_iter;
-	}
-
-	csum = bch2_checksum_bio(c, crc.csum_type, nonce, src);
-	if (bch2_crc_cmp(csum, rbio->pick.crc.csum) && !c->opts.no_data_io)
-		goto csum_err;
-
-	/*
-	 * XXX
-	 * We need to rework the narrow_crcs path to deliver the read completion
-	 * first, and then punt to a different workqueue, otherwise we're
-	 * holding up reads while doing btree updates which is bad for memory
-	 * reclaim.
-	 */
-	if (unlikely(rbio->narrow_crcs))
-		bch2_rbio_narrow_crcs(rbio);
-
-	if (rbio->flags & BCH_READ_NODECODE)
-		goto nodecode;
-
-	/* Adjust crc to point to subset of data we want: */
-	crc.offset     += rbio->offset_into_extent;
-	crc.live_size	= bvec_iter_sectors(rbio->bvec_iter);
-
-	if (crc_is_compressed(crc)) {
-		ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src);
-		if (ret)
-			goto decrypt_err;
-
-		if (bch2_bio_uncompress(c, src, dst, dst_iter, crc) &&
-		    !c->opts.no_data_io)
-			goto decompression_err;
-	} else {
-		/* don't need to decrypt the entire bio: */
-		nonce = nonce_add(nonce, crc.offset << 9);
-		bio_advance(src, crc.offset << 9);
-
-		BUG_ON(src->bi_iter.bi_size < dst_iter.bi_size);
-		src->bi_iter.bi_size = dst_iter.bi_size;
-
-		ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src);
-		if (ret)
-			goto decrypt_err;
-
-		if (rbio->bounce) {
-			struct bvec_iter src_iter = src->bi_iter;
-
-			bio_copy_data_iter(dst, &dst_iter, src, &src_iter);
-		}
-	}
-
-	if (rbio->promote) {
-		/*
-		 * Re encrypt data we decrypted, so it's consistent with
-		 * rbio->crc:
-		 */
-		ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src);
-		if (ret)
-			goto decrypt_err;
-
-		promote_start(rbio->promote, rbio);
-		rbio->promote = NULL;
-	}
-nodecode:
-	if (likely(!(rbio->flags & BCH_READ_IN_RETRY))) {
-		rbio = bch2_rbio_free(rbio);
-		bch2_rbio_done(rbio);
-	}
-out:
-	memalloc_nofs_restore(nofs_flags);
-	return;
-csum_err:
-	/*
-	 * Checksum error: if the bio wasn't bounced, we may have been
-	 * reading into buffers owned by userspace (that userspace can
-	 * scribble over) - retry the read, bouncing it this time:
-	 */
-	if (!rbio->bounce && (rbio->flags & BCH_READ_USER_MAPPED)) {
-		rbio->flags |= BCH_READ_MUST_BOUNCE;
-		bch2_rbio_error(rbio, READ_RETRY, BLK_STS_IOERR);
-		goto out;
-	}
-
-	bch_err_inum_offset_ratelimited(ca,
-		rbio->read_pos.inode,
-		rbio->read_pos.offset << 9,
-		"data checksum error: expected %0llx:%0llx got %0llx:%0llx (type %s)",
-		rbio->pick.crc.csum.hi, rbio->pick.crc.csum.lo,
-		csum.hi, csum.lo, bch2_csum_types[crc.csum_type]);
-	bch2_io_error(ca);
-	bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
-	goto out;
-decompression_err:
-	bch_err_inum_offset_ratelimited(c, rbio->read_pos.inode,
-					rbio->read_pos.offset << 9,
-					"decompression error");
-	bch2_rbio_error(rbio, READ_ERR, BLK_STS_IOERR);
-	goto out;
-decrypt_err:
-	bch_err_inum_offset_ratelimited(c, rbio->read_pos.inode,
-					rbio->read_pos.offset << 9,
-					"decrypt error");
-	bch2_rbio_error(rbio, READ_ERR, BLK_STS_IOERR);
-	goto out;
-}
-
-static void bch2_read_endio(struct bio *bio)
+void bch2_fs_io_write_exit(struct bch_fs *c)
 {
-	struct bch_read_bio *rbio =
-		container_of(bio, struct bch_read_bio, bio);
-	struct bch_fs *c	= rbio->c;
-	struct bch_dev *ca	= bch_dev_bkey_exists(c, rbio->pick.ptr.dev);
-	struct workqueue_struct *wq = NULL;
-	enum rbio_context context = RBIO_CONTEXT_NULL;
-
-	if (rbio->have_ioref) {
-		bch2_latency_acct(ca, rbio->submit_time, READ);
-		percpu_ref_put(&ca->io_ref);
-	}
-
-	if (!rbio->split)
-		rbio->bio.bi_end_io = rbio->end_io;
-
-	if (bch2_dev_inum_io_err_on(bio->bi_status, ca,
-				    rbio->read_pos.inode,
-				    rbio->read_pos.offset,
-				    "data read error: %s",
-			       bch2_blk_status_to_str(bio->bi_status))) {
-		bch2_rbio_error(rbio, READ_RETRY_AVOID, bio->bi_status);
-		return;
-	}
-
-	if (((rbio->flags & BCH_READ_RETRY_IF_STALE) && race_fault()) ||
-	    ptr_stale(ca, &rbio->pick.ptr)) {
-		trace_and_count(c, read_reuse_race, &rbio->bio);
-
-		if (rbio->flags & BCH_READ_RETRY_IF_STALE)
-			bch2_rbio_error(rbio, READ_RETRY, BLK_STS_AGAIN);
-		else
-			bch2_rbio_error(rbio, READ_ERR, BLK_STS_AGAIN);
-		return;
-	}
-
-	if (rbio->narrow_crcs ||
-	    rbio->promote ||
-	    crc_is_compressed(rbio->pick.crc) ||
-	    bch2_csum_type_is_encryption(rbio->pick.crc.csum_type))
-		context = RBIO_CONTEXT_UNBOUND,	wq = system_unbound_wq;
-	else if (rbio->pick.crc.csum_type)
-		context = RBIO_CONTEXT_HIGHPRI,	wq = system_highpri_wq;
-
-	bch2_rbio_punt(rbio, __bch2_read_endio, context, wq);
-}
-
-int __bch2_read_indirect_extent(struct btree_trans *trans,
-				unsigned *offset_into_extent,
-				struct bkey_buf *orig_k)
-{
-	struct btree_iter iter;
-	struct bkey_s_c k;
-	u64 reflink_offset;
-	int ret;
-
-	reflink_offset = le64_to_cpu(bkey_i_to_reflink_p(orig_k->k)->v.idx) +
-		*offset_into_extent;
-
-	k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_reflink,
-			       POS(0, reflink_offset), 0);
-	ret = bkey_err(k);
-	if (ret)
-		goto err;
-
-	if (k.k->type != KEY_TYPE_reflink_v &&
-	    k.k->type != KEY_TYPE_indirect_inline_data) {
-		bch_err_inum_offset_ratelimited(trans->c,
-			orig_k->k->k.p.inode,
-			orig_k->k->k.p.offset << 9,
-			"%llu len %u points to nonexistent indirect extent %llu",
-			orig_k->k->k.p.offset,
-			orig_k->k->k.size,
-			reflink_offset);
-		bch2_inconsistent_error(trans->c);
-		ret = -EIO;
-		goto err;
-	}
-
-	*offset_into_extent = iter.pos.offset - bkey_start_offset(k.k);
-	bch2_bkey_buf_reassemble(orig_k, trans->c, k);
-err:
-	bch2_trans_iter_exit(trans, &iter);
-	return ret;
-}
-
-static noinline void read_from_stale_dirty_pointer(struct btree_trans *trans,
-						   struct bkey_s_c k,
-						   struct bch_extent_ptr ptr)
-{
-	struct bch_fs *c = trans->c;
-	struct bch_dev *ca = bch_dev_bkey_exists(c, ptr.dev);
-	struct btree_iter iter;
-	struct printbuf buf = PRINTBUF;
-	int ret;
-
-	bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
-			     PTR_BUCKET_POS(c, &ptr),
-			     BTREE_ITER_CACHED);
-
-	prt_printf(&buf, "Attempting to read from stale dirty pointer:");
-	printbuf_indent_add(&buf, 2);
-	prt_newline(&buf);
-
-	bch2_bkey_val_to_text(&buf, c, k);
-	prt_newline(&buf);
-
-	prt_printf(&buf, "memory gen: %u", *bucket_gen(ca, iter.pos.offset));
-
-	ret = lockrestart_do(trans, bkey_err(k = bch2_btree_iter_peek_slot(&iter)));
-	if (!ret) {
-		prt_newline(&buf);
-		bch2_bkey_val_to_text(&buf, c, k);
-	}
-
-	bch2_fs_inconsistent(c, "%s", buf.buf);
-
-	bch2_trans_iter_exit(trans, &iter);
-	printbuf_exit(&buf);
-}
-
-int __bch2_read_extent(struct btree_trans *trans, struct bch_read_bio *orig,
-		       struct bvec_iter iter, struct bpos read_pos,
-		       enum btree_id data_btree, struct bkey_s_c k,
-		       unsigned offset_into_extent,
-		       struct bch_io_failures *failed, unsigned flags)
-{
-	struct bch_fs *c = trans->c;
-	struct extent_ptr_decoded pick;
-	struct bch_read_bio *rbio = NULL;
-	struct bch_dev *ca = NULL;
-	struct promote_op *promote = NULL;
-	bool bounce = false, read_full = false, narrow_crcs = false;
-	struct bpos data_pos = bkey_start_pos(k.k);
-	int pick_ret;
-
-	if (bkey_extent_is_inline_data(k.k)) {
-		unsigned bytes = min_t(unsigned, iter.bi_size,
-				       bkey_inline_data_bytes(k.k));
-
-		swap(iter.bi_size, bytes);
-		memcpy_to_bio(&orig->bio, iter, bkey_inline_data_p(k));
-		swap(iter.bi_size, bytes);
-		bio_advance_iter(&orig->bio, &iter, bytes);
-		zero_fill_bio_iter(&orig->bio, iter);
-		goto out_read_done;
-	}
-retry_pick:
-	pick_ret = bch2_bkey_pick_read_device(c, k, failed, &pick);
-
-	/* hole or reservation - just zero fill: */
-	if (!pick_ret)
-		goto hole;
-
-	if (pick_ret < 0) {
-		bch_err_inum_offset_ratelimited(c,
-				read_pos.inode, read_pos.offset << 9,
-				"no device to read from");
-		goto err;
-	}
-
-	ca = bch_dev_bkey_exists(c, pick.ptr.dev);
-
-	/*
-	 * Stale dirty pointers are treated as IO errors, but @failed isn't
-	 * allocated unless we're in the retry path - so if we're not in the
-	 * retry path, don't check here, it'll be caught in bch2_read_endio()
-	 * and we'll end up in the retry path:
-	 */
-	if ((flags & BCH_READ_IN_RETRY) &&
-	    !pick.ptr.cached &&
-	    unlikely(ptr_stale(ca, &pick.ptr))) {
-		read_from_stale_dirty_pointer(trans, k, pick.ptr);
-		bch2_mark_io_failure(failed, &pick);
-		goto retry_pick;
-	}
-
-	/*
-	 * Unlock the iterator while the btree node's lock is still in
-	 * cache, before doing the IO:
-	 */
-	bch2_trans_unlock(trans);
-
-	if (flags & BCH_READ_NODECODE) {
-		/*
-		 * can happen if we retry, and the extent we were going to read
-		 * has been merged in the meantime:
-		 */
-		if (pick.crc.compressed_size > orig->bio.bi_vcnt * PAGE_SECTORS)
-			goto hole;
-
-		iter.bi_size	= pick.crc.compressed_size << 9;
-		goto get_bio;
-	}
-
-	if (!(flags & BCH_READ_LAST_FRAGMENT) ||
-	    bio_flagged(&orig->bio, BIO_CHAIN))
-		flags |= BCH_READ_MUST_CLONE;
-
-	narrow_crcs = !(flags & BCH_READ_IN_RETRY) &&
-		bch2_can_narrow_extent_crcs(k, pick.crc);
-
-	if (narrow_crcs && (flags & BCH_READ_USER_MAPPED))
-		flags |= BCH_READ_MUST_BOUNCE;
-
-	EBUG_ON(offset_into_extent + bvec_iter_sectors(iter) > k.k->size);
-
-	if (crc_is_compressed(pick.crc) ||
-	    (pick.crc.csum_type != BCH_CSUM_none &&
-	     (bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
-	      (bch2_csum_type_is_encryption(pick.crc.csum_type) &&
-	       (flags & BCH_READ_USER_MAPPED)) ||
-	      (flags & BCH_READ_MUST_BOUNCE)))) {
-		read_full = true;
-		bounce = true;
-	}
-
-	if (orig->opts.promote_target)
-		promote = promote_alloc(trans, iter, k, &pick, orig->opts, flags,
-					&rbio, &bounce, &read_full);
-
-	if (!read_full) {
-		EBUG_ON(crc_is_compressed(pick.crc));
-		EBUG_ON(pick.crc.csum_type &&
-			(bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
-			 bvec_iter_sectors(iter) != pick.crc.live_size ||
-			 pick.crc.offset ||
-			 offset_into_extent));
-
-		data_pos.offset += offset_into_extent;
-		pick.ptr.offset += pick.crc.offset +
-			offset_into_extent;
-		offset_into_extent		= 0;
-		pick.crc.compressed_size	= bvec_iter_sectors(iter);
-		pick.crc.uncompressed_size	= bvec_iter_sectors(iter);
-		pick.crc.offset			= 0;
-		pick.crc.live_size		= bvec_iter_sectors(iter);
-		offset_into_extent		= 0;
-	}
-get_bio:
-	if (rbio) {
-		/*
-		 * promote already allocated bounce rbio:
-		 * promote needs to allocate a bio big enough for uncompressing
-		 * data in the write path, but we're not going to use it all
-		 * here:
-		 */
-		EBUG_ON(rbio->bio.bi_iter.bi_size <
-		       pick.crc.compressed_size << 9);
-		rbio->bio.bi_iter.bi_size =
-			pick.crc.compressed_size << 9;
-	} else if (bounce) {
-		unsigned sectors = pick.crc.compressed_size;
-
-		rbio = rbio_init(bio_alloc_bioset(NULL,
-						  DIV_ROUND_UP(sectors, PAGE_SECTORS),
-						  0,
-						  GFP_NOFS,
-						  &c->bio_read_split),
-				 orig->opts);
-
-		bch2_bio_alloc_pages_pool(c, &rbio->bio, sectors << 9);
-		rbio->bounce	= true;
-		rbio->split	= true;
-	} else if (flags & BCH_READ_MUST_CLONE) {
-		/*
-		 * Have to clone if there were any splits, due to error
-		 * reporting issues (if a split errored, and retrying didn't
-		 * work, when it reports the error to its parent (us) we don't
-		 * know if the error was from our bio, and we should retry, or
-		 * from the whole bio, in which case we don't want to retry and
-		 * lose the error)
-		 */
-		rbio = rbio_init(bio_alloc_clone(NULL, &orig->bio, GFP_NOFS,
-						 &c->bio_read_split),
-				 orig->opts);
-		rbio->bio.bi_iter = iter;
-		rbio->split	= true;
-	} else {
-		rbio = orig;
-		rbio->bio.bi_iter = iter;
-		EBUG_ON(bio_flagged(&rbio->bio, BIO_CHAIN));
-	}
-
-	EBUG_ON(bio_sectors(&rbio->bio) != pick.crc.compressed_size);
-
-	rbio->c			= c;
-	rbio->submit_time	= local_clock();
-	if (rbio->split)
-		rbio->parent	= orig;
-	else
-		rbio->end_io	= orig->bio.bi_end_io;
-	rbio->bvec_iter		= iter;
-	rbio->offset_into_extent= offset_into_extent;
-	rbio->flags		= flags;
-	rbio->have_ioref	= pick_ret > 0 && bch2_dev_get_ioref(ca, READ);
-	rbio->narrow_crcs	= narrow_crcs;
-	rbio->hole		= 0;
-	rbio->retry		= 0;
-	rbio->context		= 0;
-	/* XXX: only initialize this if needed */
-	rbio->devs_have		= bch2_bkey_devs(k);
-	rbio->pick		= pick;
-	rbio->subvol		= orig->subvol;
-	rbio->read_pos		= read_pos;
-	rbio->data_btree	= data_btree;
-	rbio->data_pos		= data_pos;
-	rbio->version		= k.k->version;
-	rbio->promote		= promote;
-	INIT_WORK(&rbio->work, NULL);
-
-	rbio->bio.bi_opf	= orig->bio.bi_opf;
-	rbio->bio.bi_iter.bi_sector = pick.ptr.offset;
-	rbio->bio.bi_end_io	= bch2_read_endio;
-
-	if (rbio->bounce)
-		trace_and_count(c, read_bounce, &rbio->bio);
-
-	this_cpu_add(c->counters[BCH_COUNTER_io_read], bio_sectors(&rbio->bio));
-	bch2_increment_clock(c, bio_sectors(&rbio->bio), READ);
-
-	/*
-	 * If it's being moved internally, we don't want to flag it as a cache
-	 * hit:
-	 */
-	if (pick.ptr.cached && !(flags & BCH_READ_NODECODE))
-		bch2_bucket_io_time_reset(trans, pick.ptr.dev,
-			PTR_BUCKET_NR(ca, &pick.ptr), READ);
-
-	if (!(flags & (BCH_READ_IN_RETRY|BCH_READ_LAST_FRAGMENT))) {
-		bio_inc_remaining(&orig->bio);
-		trace_and_count(c, read_split, &orig->bio);
-	}
-
-	if (!rbio->pick.idx) {
-		if (!rbio->have_ioref) {
-			bch_err_inum_offset_ratelimited(c,
-					read_pos.inode,
-					read_pos.offset << 9,
-					"no device to read from");
-			bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
-			goto out;
-		}
-
-		this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_user],
-			     bio_sectors(&rbio->bio));
-		bio_set_dev(&rbio->bio, ca->disk_sb.bdev);
-
-		if (unlikely(c->opts.no_data_io)) {
-			if (likely(!(flags & BCH_READ_IN_RETRY)))
-				bio_endio(&rbio->bio);
-		} else {
-			if (likely(!(flags & BCH_READ_IN_RETRY)))
-				submit_bio(&rbio->bio);
-			else
-				submit_bio_wait(&rbio->bio);
-		}
-
-		/*
-		 * We just submitted IO which may block, we expect relock fail
-		 * events and shouldn't count them:
-		 */
-		trans->notrace_relock_fail = true;
-	} else {
-		/* Attempting reconstruct read: */
-		if (bch2_ec_read_extent(c, rbio)) {
-			bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
-			goto out;
-		}
-
-		if (likely(!(flags & BCH_READ_IN_RETRY)))
-			bio_endio(&rbio->bio);
-	}
-out:
-	if (likely(!(flags & BCH_READ_IN_RETRY))) {
-		return 0;
-	} else {
-		int ret;
-
-		rbio->context = RBIO_CONTEXT_UNBOUND;
-		bch2_read_endio(&rbio->bio);
-
-		ret = rbio->retry;
-		rbio = bch2_rbio_free(rbio);
-
-		if (ret == READ_RETRY_AVOID) {
-			bch2_mark_io_failure(failed, &pick);
-			ret = READ_RETRY;
-		}
-
-		if (!ret)
-			goto out_read_done;
-
-		return ret;
-	}
-
-err:
-	if (flags & BCH_READ_IN_RETRY)
-		return READ_ERR;
-
-	orig->bio.bi_status = BLK_STS_IOERR;
-	goto out_read_done;
-
-hole:
-	/*
-	 * won't normally happen in the BCH_READ_NODECODE
-	 * (bch2_move_extent()) path, but if we retry and the extent we wanted
-	 * to read no longer exists we have to signal that:
-	 */
-	if (flags & BCH_READ_NODECODE)
-		orig->hole = true;
-
-	zero_fill_bio_iter(&orig->bio, iter);
-out_read_done:
-	if (flags & BCH_READ_LAST_FRAGMENT)
-		bch2_rbio_done(orig);
-	return 0;
-}
-
-void __bch2_read(struct bch_fs *c, struct bch_read_bio *rbio,
-		 struct bvec_iter bvec_iter, subvol_inum inum,
-		 struct bch_io_failures *failed, unsigned flags)
-{
-	struct btree_trans trans;
-	struct btree_iter iter;
-	struct bkey_buf sk;
-	struct bkey_s_c k;
-	u32 snapshot;
-	int ret;
-
-	BUG_ON(flags & BCH_READ_NODECODE);
-
-	bch2_bkey_buf_init(&sk);
-	bch2_trans_init(&trans, c, 0, 0);
-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, bvec_iter.bi_sector, snapshot),
-			     BTREE_ITER_SLOTS);
-	while (1) {
-		unsigned bytes, sectors, offset_into_extent;
-		enum btree_id data_btree = BTREE_ID_extents;
-
-		/*
-		 * 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;
-
-		bch2_btree_iter_set_pos(&iter,
-				POS(inum.inum, bvec_iter.bi_sector));
-
-		k = bch2_btree_iter_peek_slot(&iter);
-		ret = bkey_err(k);
-		if (ret)
-			break;
-
-		offset_into_extent = iter.pos.offset -
-			bkey_start_offset(k.k);
-		sectors = k.k->size - offset_into_extent;
-
-		bch2_bkey_buf_reassemble(&sk, c, k);
-
-		ret = bch2_read_indirect_extent(&trans, &data_btree,
-					&offset_into_extent, &sk);
-		if (ret)
-			break;
-
-		k = bkey_i_to_s_c(sk.k);
-
-		/*
-		 * With indirect extents, the amount of data to read is the min
-		 * of the original extent and the indirect extent:
-		 */
-		sectors = min(sectors, k.k->size - offset_into_extent);
-
-		bytes = min(sectors, bvec_iter_sectors(bvec_iter)) << 9;
-		swap(bvec_iter.bi_size, bytes);
-
-		if (bvec_iter.bi_size == bytes)
-			flags |= BCH_READ_LAST_FRAGMENT;
-
-		ret = __bch2_read_extent(&trans, rbio, bvec_iter, iter.pos,
-					 data_btree, k,
-					 offset_into_extent, failed, flags);
-		if (ret)
-			break;
-
-		if (flags & BCH_READ_LAST_FRAGMENT)
-			break;
-
-		swap(bvec_iter.bi_size, bytes);
-		bio_advance_iter(&rbio->bio, &bvec_iter, 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) ||
-	    ret == READ_RETRY ||
-	    ret == READ_RETRY_AVOID)
-		goto retry;
-
-	bch2_trans_exit(&trans);
-	bch2_bkey_buf_exit(&sk, c);
-
-	if (ret) {
-		bch_err_inum_offset_ratelimited(c, inum.inum,
-						bvec_iter.bi_sector << 9,
-						"read error %i from btree lookup", ret);
-		rbio->bio.bi_status = BLK_STS_IOERR;
-		bch2_rbio_done(rbio);
-	}
-}
-
-void bch2_fs_io_exit(struct bch_fs *c)
-{
-	if (c->promote_table.tbl)
-		rhashtable_destroy(&c->promote_table);
 	mempool_exit(&c->bio_bounce_pages);
 	bioset_exit(&c->bio_write);
-	bioset_exit(&c->bio_read_split);
-	bioset_exit(&c->bio_read);
 }
 
-int bch2_fs_io_init(struct bch_fs *c)
+int bch2_fs_io_write_init(struct bch_fs *c)
 {
-	if (bioset_init(&c->bio_read, 1, offsetof(struct bch_read_bio, bio),
-			BIOSET_NEED_BVECS))
-		return -BCH_ERR_ENOMEM_bio_read_init;
-
-	if (bioset_init(&c->bio_read_split, 1, offsetof(struct bch_read_bio, bio),
-			BIOSET_NEED_BVECS))
-		return -BCH_ERR_ENOMEM_bio_read_split_init;
-
 	if (bioset_init(&c->bio_write, 1, offsetof(struct bch_write_bio, bio),
 			BIOSET_NEED_BVECS))
 		return -BCH_ERR_ENOMEM_bio_write_init;
@@ -3044,8 +1666,5 @@ int bch2_fs_io_init(struct bch_fs *c)
 				   PAGE_SIZE, 0))
 		return -BCH_ERR_ENOMEM_bio_bounce_pages_init;
 
-	if (rhashtable_init(&c->promote_table, &bch_promote_params))
-		return -BCH_ERR_ENOMEM_promote_table_init;
-
 	return 0;
 }
diff --git a/fs/bcachefs/io_write.h b/fs/bcachefs/io_write.h
new file mode 100644
index 000000000000..9323167229ee
--- /dev/null
+++ b/fs/bcachefs/io_write.h
@@ -0,0 +1,110 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _BCACHEFS_IO_WRITE_H
+#define _BCACHEFS_IO_WRITE_H
+
+#include "checksum.h"
+#include "io_write_types.h"
+
+#define to_wbio(_bio)			\
+	container_of((_bio), struct bch_write_bio, bio)
+
+void bch2_bio_free_pages_pool(struct bch_fs *, struct bio *);
+void bch2_bio_alloc_pages_pool(struct bch_fs *, struct bio *, size_t);
+
+#ifndef CONFIG_BCACHEFS_NO_LATENCY_ACCT
+void bch2_latency_acct(struct bch_dev *, u64, int);
+#else
+static inline void bch2_latency_acct(struct bch_dev *ca, u64 submit_time, int rw) {}
+#endif
+
+void bch2_submit_wbio_replicas(struct bch_write_bio *, struct bch_fs *,
+			       enum bch_data_type, const struct bkey_i *, bool);
+
+#define BCH_WRITE_FLAGS()		\
+	x(ALLOC_NOWAIT)			\
+	x(CACHED)			\
+	x(DATA_ENCODED)			\
+	x(PAGES_STABLE)			\
+	x(PAGES_OWNED)			\
+	x(ONLY_SPECIFIED_DEVS)		\
+	x(WROTE_DATA_INLINE)		\
+	x(FROM_INTERNAL)		\
+	x(CHECK_ENOSPC)			\
+	x(SYNC)				\
+	x(MOVE)				\
+	x(IN_WORKER)			\
+	x(DONE)				\
+	x(IO_ERROR)			\
+	x(CONVERT_UNWRITTEN)
+
+enum __bch_write_flags {
+#define x(f)	__BCH_WRITE_##f,
+	BCH_WRITE_FLAGS()
+#undef x
+};
+
+enum bch_write_flags {
+#define x(f)	BCH_WRITE_##f = BIT(__BCH_WRITE_##f),
+	BCH_WRITE_FLAGS()
+#undef x
+};
+
+static inline struct workqueue_struct *index_update_wq(struct bch_write_op *op)
+{
+	return op->watermark == BCH_WATERMARK_copygc
+		? op->c->copygc_wq
+		: op->c->btree_update_wq;
+}
+
+int bch2_sum_sector_overwrites(struct btree_trans *, struct btree_iter *,
+			       struct bkey_i *, bool *, s64 *, s64 *);
+int bch2_extent_update(struct btree_trans *, subvol_inum,
+		       struct btree_iter *, struct bkey_i *,
+		       struct disk_reservation *, u64, s64 *, bool);
+
+static inline void bch2_write_op_init(struct bch_write_op *op, struct bch_fs *c,
+				      struct bch_io_opts opts)
+{
+	op->c			= c;
+	op->end_io		= NULL;
+	op->flags		= 0;
+	op->written		= 0;
+	op->error		= 0;
+	op->csum_type		= bch2_data_checksum_type(c, opts);
+	op->compression_opt	= opts.compression;
+	op->nr_replicas		= 0;
+	op->nr_replicas_required = c->opts.data_replicas_required;
+	op->watermark		= BCH_WATERMARK_normal;
+	op->incompressible	= 0;
+	op->open_buckets.nr	= 0;
+	op->devs_have.nr	= 0;
+	op->target		= 0;
+	op->opts		= opts;
+	op->subvol		= 0;
+	op->pos			= POS_MAX;
+	op->version		= ZERO_VERSION;
+	op->write_point		= (struct write_point_specifier) { 0 };
+	op->res			= (struct disk_reservation) { 0 };
+	op->new_i_size		= U64_MAX;
+	op->i_sectors_delta	= 0;
+	op->devs_need_flush	= NULL;
+}
+
+void bch2_write(struct closure *);
+
+void bch2_write_point_do_index_updates(struct work_struct *);
+
+static inline struct bch_write_bio *wbio_init(struct bio *bio)
+{
+	struct bch_write_bio *wbio = to_wbio(bio);
+
+	memset(&wbio->wbio, 0, sizeof(wbio->wbio));
+	return wbio;
+}
+
+void bch2_write_op_to_text(struct printbuf *, struct bch_write_op *);
+
+void bch2_fs_io_write_exit(struct bch_fs *);
+int bch2_fs_io_write_init(struct bch_fs *);
+
+#endif /* _BCACHEFS_IO_WRITE_H */
diff --git a/fs/bcachefs/io_types.h b/fs/bcachefs/io_write_types.h
index 737f16d78c48..c7f97c2c4805 100644
--- a/fs/bcachefs/io_types.h
+++ b/fs/bcachefs/io_write_types.h
@@ -1,6 +1,6 @@
 /* SPDX-License-Identifier: GPL-2.0 */
-#ifndef _BCACHEFS_IO_TYPES_H
-#define _BCACHEFS_IO_TYPES_H
+#ifndef _BCACHEFS_IO_WRITE_TYPES_H
+#define _BCACHEFS_IO_WRITE_TYPES_H
 
 #include "alloc_types.h"
 #include "btree_types.h"
@@ -13,75 +13,6 @@
 #include <linux/llist.h>
 #include <linux/workqueue.h>
 
-struct bch_read_bio {
-	struct bch_fs		*c;
-	u64			start_time;
-	u64			submit_time;
-
-	/*
-	 * Reads will often have to be split, and if the extent being read from
-	 * was checksummed or compressed we'll also have to allocate bounce
-	 * buffers and copy the data back into the original bio.
-	 *
-	 * If we didn't have to split, we have to save and restore the original
-	 * bi_end_io - @split below indicates which:
-	 */
-	union {
-	struct bch_read_bio	*parent;
-	bio_end_io_t		*end_io;
-	};
-
-	/*
-	 * Saved copy of bio->bi_iter, from submission time - allows us to
-	 * resubmit on IO error, and also to copy data back to the original bio
-	 * when we're bouncing:
-	 */
-	struct bvec_iter	bvec_iter;
-
-	unsigned		offset_into_extent;
-
-	u16			flags;
-	union {
-	struct {
-	u16			bounce:1,
-				split:1,
-				kmalloc:1,
-				have_ioref:1,
-				narrow_crcs:1,
-				hole:1,
-				retry:2,
-				context:2;
-	};
-	u16			_state;
-	};
-
-	struct bch_devs_list	devs_have;
-
-	struct extent_ptr_decoded pick;
-
-	/*
-	 * pos we read from - different from data_pos for indirect extents:
-	 */
-	u32			subvol;
-	struct bpos		read_pos;
-
-	/*
-	 * start pos of data we read (may not be pos of data we want) - for
-	 * promote, narrow extents paths:
-	 */
-	enum btree_id		data_btree;
-	struct bpos		data_pos;
-	struct bversion		version;
-
-	struct promote_op	*promote;
-
-	struct bch_io_opts	opts;
-
-	struct work_struct	work;
-
-	struct bio		bio;
-};
-
 struct bch_write_bio {
 	struct_group(wbio,
 	struct bch_fs		*c;
@@ -162,4 +93,4 @@ struct bch_write_op {
 	struct bch_write_bio	wbio;
 };
 
-#endif /* _BCACHEFS_IO_TYPES_H */
+#endif /* _BCACHEFS_IO_WRITE_TYPES_H */
diff --git a/fs/bcachefs/journal_io.c b/fs/bcachefs/journal_io.c
index 34740dca4b15..0e606009dc46 100644
--- a/fs/bcachefs/journal_io.c
+++ b/fs/bcachefs/journal_io.c
@@ -8,7 +8,6 @@
 #include "checksum.h"
 #include "disk_groups.h"
 #include "error.h"
-#include "io.h"
 #include "journal.h"
 #include "journal_io.h"
 #include "journal_reclaim.h"
diff --git a/fs/bcachefs/migrate.c b/fs/bcachefs/migrate.c
index 3d7c5b919421..4746dfa7af97 100644
--- a/fs/bcachefs/migrate.c
+++ b/fs/bcachefs/migrate.c
@@ -10,7 +10,7 @@
 #include "buckets.h"
 #include "errcode.h"
 #include "extents.h"
-#include "io.h"
+#include "io_write.h"
 #include "journal.h"
 #include "keylist.h"
 #include "migrate.h"
diff --git a/fs/bcachefs/move.c b/fs/bcachefs/move.c
index fb76a1dac74e..ac4df53bfde2 100644
--- a/fs/bcachefs/move.c
+++ b/fs/bcachefs/move.c
@@ -14,7 +14,8 @@
 #include "errcode.h"
 #include "error.h"
 #include "inode.h"
-#include "io.h"
+#include "io_read.h"
+#include "io_write.h"
 #include "journal_reclaim.h"
 #include "keylist.h"
 #include "move.h"
diff --git a/fs/bcachefs/move.h b/fs/bcachefs/move.h
index c3136abe8587..cbdd58db8782 100644
--- a/fs/bcachefs/move.h
+++ b/fs/bcachefs/move.h
@@ -2,6 +2,7 @@
 #ifndef _BCACHEFS_MOVE_H
 #define _BCACHEFS_MOVE_H
 
+#include "bcachefs_ioctl.h"
 #include "btree_iter.h"
 #include "buckets.h"
 #include "data_update.h"
diff --git a/fs/bcachefs/movinggc.c b/fs/bcachefs/movinggc.c
index ac658e99bf57..2371fd61ea58 100644
--- a/fs/bcachefs/movinggc.c
+++ b/fs/bcachefs/movinggc.c
@@ -13,25 +13,17 @@
 #include "btree_write_buffer.h"
 #include "buckets.h"
 #include "clock.h"
-#include "disk_groups.h"
 #include "errcode.h"
 #include "error.h"
-#include "extents.h"
-#include "eytzinger.h"
-#include "io.h"
-#include "keylist.h"
 #include "lru.h"
 #include "move.h"
 #include "movinggc.h"
-#include "super-io.h"
 #include "trace.h"
 
-#include <linux/bsearch.h>
 #include <linux/freezer.h>
 #include <linux/kthread.h>
 #include <linux/math64.h>
 #include <linux/sched/task.h>
-#include <linux/sort.h>
 #include <linux/wait.h>
 
 struct buckets_in_flight {
diff --git a/fs/bcachefs/rebalance.c b/fs/bcachefs/rebalance.c
index 016cf0834b3d..568f1e8e7507 100644
--- a/fs/bcachefs/rebalance.c
+++ b/fs/bcachefs/rebalance.c
@@ -8,8 +8,6 @@
 #include "compress.h"
 #include "disk_groups.h"
 #include "errcode.h"
-#include "extents.h"
-#include "io.h"
 #include "move.h"
 #include "rebalance.h"
 #include "super-io.h"
diff --git a/fs/bcachefs/reflink.c b/fs/bcachefs/reflink.c
index 39f711d5069e..f155428ff395 100644
--- a/fs/bcachefs/reflink.c
+++ b/fs/bcachefs/reflink.c
@@ -5,9 +5,11 @@
 #include "buckets.h"
 #include "extents.h"
 #include "inode.h"
-#include "io.h"
+#include "io_misc.h"
+#include "io_write.h"
 #include "reflink.h"
 #include "subvolume.h"
+#include "super-io.h"
 
 #include <linux/sched/signal.h>
 
diff --git a/fs/bcachefs/super-io.c b/fs/bcachefs/super-io.c
index f01883e785a5..5a1115396edc 100644
--- a/fs/bcachefs/super-io.c
+++ b/fs/bcachefs/super-io.c
@@ -6,7 +6,6 @@
 #include "disk_groups.h"
 #include "ec.h"
 #include "error.h"
-#include "io.h"
 #include "journal.h"
 #include "journal_sb.h"
 #include "journal_seq_blacklist.h"
diff --git a/fs/bcachefs/super.c b/fs/bcachefs/super.c
index 7cfc04947717..55176023f15b 100644
--- a/fs/bcachefs/super.c
+++ b/fs/bcachefs/super.c
@@ -35,7 +35,8 @@
 #include "fs-io-direct.h"
 #include "fsck.h"
 #include "inode.h"
-#include "io.h"
+#include "io_read.h"
+#include "io_write.h"
 #include "journal.h"
 #include "journal_reclaim.h"
 #include "journal_seq_blacklist.h"
@@ -483,7 +484,8 @@ static void __bch2_fs_free(struct bch_fs *c)
 	bch2_fs_fsio_exit(c);
 	bch2_fs_ec_exit(c);
 	bch2_fs_encryption_exit(c);
-	bch2_fs_io_exit(c);
+	bch2_fs_io_write_exit(c);
+	bch2_fs_io_read_exit(c);
 	bch2_fs_buckets_waiting_for_journal_exit(c);
 	bch2_fs_btree_interior_update_exit(c);
 	bch2_fs_btree_iter_exit(c);
@@ -848,7 +850,8 @@ static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_opts opts)
 	    bch2_fs_buckets_waiting_for_journal_init(c) ?:
 	    bch2_fs_btree_write_buffer_init(c) ?:
 	    bch2_fs_subvolumes_init(c) ?:
-	    bch2_fs_io_init(c) ?:
+	    bch2_fs_io_read_init(c) ?:
+	    bch2_fs_io_write_init(c) ?:
 	    bch2_fs_nocow_locking_init(c) ?:
 	    bch2_fs_encryption_init(c) ?:
 	    bch2_fs_compress_init(c) ?: