Commit a2cb8a62 authored by Kent Overstreet's avatar Kent Overstreet
Browse files

bcachefs: Self healing on read IO error 



This repurposes the promote path, which already knows how to call
data_update() after a read: we now automatically rewrite bad data when
we get a read error and then successfully retry from a different
replica.

Signed-off-by: default avatarKent Overstreet <kent.overstreet@linux.dev>
parent b1d63b06

fs/bcachefs/extents.c

+4 −4
Original line number Diff line number Diff line
@@ -37,7 +37,7 @@ static void bch2_extent_crc_pack(union bch_extent_crc *, 
				 struct bch_extent_crc_unpacked,

				 enum bch_extent_entry_type);



static struct bch_dev_io_failures *dev_io_failures(struct bch_io_failures *f,

struct bch_dev_io_failures *bch2_dev_io_failures(struct bch_io_failures *f,

						 unsigned dev)

{

	struct bch_dev_io_failures *i;
@@ -52,7 +52,7 @@ static struct bch_dev_io_failures *dev_io_failures(struct bch_io_failures *f, 
void bch2_mark_io_failure(struct bch_io_failures *failed,

			  struct extent_ptr_decoded *p)

{

	struct bch_dev_io_failures *f = dev_io_failures(failed, p->ptr.dev);

	struct bch_dev_io_failures *f = bch2_dev_io_failures(failed, p->ptr.dev);



	if (!f) {

		BUG_ON(failed->nr >= ARRAY_SIZE(failed->devs));
@@ -140,7 +140,7 @@ int bch2_bkey_pick_read_device(struct bch_fs *c, struct bkey_s_c k, 
		if (p.ptr.cached && (!ca || dev_ptr_stale_rcu(ca, &p.ptr)))

			continue;



		f = failed ? dev_io_failures(failed, p.ptr.dev) : NULL;

		f = failed ? bch2_dev_io_failures(failed, p.ptr.dev) : NULL;

		if (f)

			p.idx = f->nr_failed < f->nr_retries

				? f->idx

fs/bcachefs/extents.h

+2 −0
Original line number Diff line number Diff line
@@ -399,6 +399,8 @@ out: \ 


/* utility code common to all keys with pointers: */



struct bch_dev_io_failures *bch2_dev_io_failures(struct bch_io_failures *,

						 unsigned);

void bch2_mark_io_failure(struct bch_io_failures *,

			  struct extent_ptr_decoded *);

int bch2_bkey_pick_read_device(struct bch_fs *, struct bkey_s_c,

fs/bcachefs/io_read.c

+47 −22
Original line number Diff line number Diff line
@@ -93,8 +93,10 @@ static const struct rhashtable_params bch_promote_params = { 
static inline int should_promote(struct bch_fs *c, struct bkey_s_c k,

				  struct bpos pos,

				  struct bch_io_opts opts,

				  unsigned flags)

				  unsigned flags,

				  struct bch_io_failures *failed)

{

	if (!failed) {

		BUG_ON(!opts.promote_target);



		if (!(flags & BCH_READ_MAY_PROMOTE))
@@ -108,6 +110,7 @@ static inline int should_promote(struct bch_fs *c, struct bkey_s_c k, 


		if (bch2_target_congested(c, opts.promote_target))

			return -BCH_ERR_nopromote_congested;

	}



	if (rhashtable_lookup_fast(&c->promote_table, &pos,

				   bch_promote_params))
@@ -164,7 +167,8 @@ static struct promote_op *__promote_alloc(struct btree_trans *trans, 
					  struct extent_ptr_decoded *pick,

					  struct bch_io_opts opts,

					  unsigned sectors,

					  struct bch_read_bio **rbio)

					  struct bch_read_bio **rbio,

					  struct bch_io_failures *failed)

{

	struct bch_fs *c = trans->c;

	struct promote_op *op = NULL;
@@ -217,14 +221,28 @@ static struct promote_op *__promote_alloc(struct btree_trans *trans, 
	bio = &op->write.op.wbio.bio;

	bio_init(bio, NULL, bio->bi_inline_vecs, pages, 0);



	struct data_update_opts update_opts = {};



	if (!failed) {

		update_opts.target = opts.promote_target;

		update_opts.extra_replicas = 1;

		update_opts.write_flags = BCH_WRITE_ALLOC_NOWAIT|BCH_WRITE_CACHED;

	} else {

		update_opts.target = opts.foreground_target;



		struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);

		unsigned i = 0;

		bkey_for_each_ptr(ptrs, ptr) {

			if (bch2_dev_io_failures(failed, ptr->dev))

				update_opts.rewrite_ptrs |= BIT(i);

			i++;

		}

	}



	ret = bch2_data_update_init(trans, NULL, 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,

			},

			update_opts,

			btree_id, k);

	/*

	 * possible errors: -BCH_ERR_nocow_lock_blocked,
@@ -258,10 +276,17 @@ static struct promote_op *promote_alloc(struct btree_trans *trans, 
					unsigned flags,

					struct bch_read_bio **rbio,

					bool *bounce,

					bool *read_full)

					bool *read_full,

					struct bch_io_failures *failed)

{

	struct bch_fs *c = trans->c;

	bool promote_full = *read_full || READ_ONCE(c->promote_whole_extents);

	/*

	 * if failed != NULL we're not actually doing a promote, we're

	 * recovering from an io/checksum error

	 */

	bool promote_full = (failed ||

			     *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)
@@ -272,7 +297,7 @@ static struct promote_op *promote_alloc(struct btree_trans *trans, 
	struct promote_op *promote;

	int ret;



	ret = should_promote(c, k, pos, opts, flags);

	ret = should_promote(c, k, pos, opts, flags, failed);

	if (ret)

		goto nopromote;
@@ -280,7 +305,7 @@ static struct promote_op *promote_alloc(struct btree_trans *trans, 
				  k.k->type == KEY_TYPE_reflink_v

				  ? BTREE_ID_reflink

				  : BTREE_ID_extents,

				  k, pos, pick, opts, sectors, rbio);

				  k, pos, pick, opts, sectors, rbio, failed);

	ret = PTR_ERR_OR_ZERO(promote);

	if (ret)

		goto nopromote;
@@ -910,9 +935,9 @@ int __bch2_read_extent(struct btree_trans *trans, struct bch_read_bio *orig, 
		bounce = true;

	}



	if (orig->opts.promote_target)

	if (orig->opts.promote_target)// || failed)

		promote = promote_alloc(trans, iter, k, &pick, orig->opts, flags,

					&rbio, &bounce, &read_full);

					&rbio, &bounce, &read_full, failed);



	if (!read_full) {

		EBUG_ON(crc_is_compressed(pick.crc));