X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libbcachefs%2Fextents.c;h=82ec056f4cdbb1f4e4234fce274939b61b7a5015;hb=1f79cf3825e94fcb146d417b6dda9b94c93c7a53;hp=f66640c2a5edd73ad8c17059caed61df37a241b4;hpb=2b8c1bb0910534e8687ea3e5abf6d8bbba758247;p=bcachefs-tools-debian

diff --git a/libbcachefs/extents.c b/libbcachefs/extents.c
index f66640c..82ec056 100644
--- a/libbcachefs/extents.c
+++ b/libbcachefs/extents.c
@@ -13,6 +13,7 @@
 #include "btree_iter.h"
 #include "buckets.h"
 #include "checksum.h"
+#include "compress.h"
 #include "debug.h"
 #include "disk_groups.h"
 #include "error.h"
@@ -22,10 +23,9 @@
 #include "replicas.h"
 #include "super.h"
 #include "super-io.h"
+#include "trace.h"
 #include "util.h"
 
-#include <trace/events/bcachefs.h>
-
 static unsigned bch2_crc_field_size_max[] = {
 	[BCH_EXTENT_ENTRY_crc32] = CRC32_SIZE_MAX,
 	[BCH_EXTENT_ENTRY_crc64] = CRC64_SIZE_MAX,
@@ -115,6 +115,13 @@ int bch2_bkey_pick_read_device(struct bch_fs *c, struct bkey_s_c k,
 		return -EIO;
 
 	bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
+		/*
+		 * Unwritten extent: no need to actually read, treat it as a
+		 * hole and return 0s:
+		 */
+		if (p.ptr.unwritten)
+			return 0;
+
 		ca = bch_dev_bkey_exists(c, p.ptr.dev);
 
 		/*
@@ -156,12 +163,19 @@ int bch2_bkey_pick_read_device(struct bch_fs *c, struct bkey_s_c k,
 
 /* KEY_TYPE_btree_ptr: */
 
-const char *bch2_btree_ptr_invalid(const struct bch_fs *c, struct bkey_s_c k)
+int bch2_btree_ptr_invalid(struct bch_fs *c, struct bkey_s_c k,
+			   enum bkey_invalid_flags flags,
+			   struct printbuf *err)
 {
-	if (bkey_val_u64s(k.k) > BCH_REPLICAS_MAX)
-		return "value too big";
+	int ret = 0;
+
+	bkey_fsck_err_on(bkey_val_u64s(k.k) > BCH_REPLICAS_MAX, c, err,
+			 btree_ptr_val_too_big,
+			 "value too big (%zu > %u)", bkey_val_u64s(k.k), BCH_REPLICAS_MAX);
 
-	return bch2_bkey_ptrs_invalid(c, k);
+	ret = bch2_bkey_ptrs_invalid(c, k, flags, err);
+fsck_err:
+	return ret;
 }
 
 void bch2_btree_ptr_to_text(struct printbuf *out, struct bch_fs *c,
@@ -170,35 +184,34 @@ void bch2_btree_ptr_to_text(struct printbuf *out, struct bch_fs *c,
 	bch2_bkey_ptrs_to_text(out, c, k);
 }
 
-const char *bch2_btree_ptr_v2_invalid(const struct bch_fs *c, struct bkey_s_c k)
+int bch2_btree_ptr_v2_invalid(struct bch_fs *c, struct bkey_s_c k,
+			      enum bkey_invalid_flags flags,
+			      struct printbuf *err)
 {
-	struct bkey_s_c_btree_ptr_v2 bp = bkey_s_c_to_btree_ptr_v2(k);
-
-	if (bkey_val_bytes(k.k) <= sizeof(*bp.v))
-		return "value too small";
-
-	if (bkey_val_u64s(k.k) > BKEY_BTREE_PTR_VAL_U64s_MAX)
-		return "value too big";
+	int ret = 0;
 
-	if (c->sb.version < bcachefs_metadata_version_snapshot &&
-	    bp.v->min_key.snapshot)
-		return "invalid min_key.snapshot";
+	bkey_fsck_err_on(bkey_val_u64s(k.k) > BKEY_BTREE_PTR_VAL_U64s_MAX, c, err,
+			 btree_ptr_v2_val_too_big,
+			 "value too big (%zu > %zu)",
+			 bkey_val_u64s(k.k), BKEY_BTREE_PTR_VAL_U64s_MAX);
 
-	return bch2_bkey_ptrs_invalid(c, k);
+	ret = bch2_bkey_ptrs_invalid(c, k, flags, err);
+fsck_err:
+	return ret;
 }
 
 void bch2_btree_ptr_v2_to_text(struct printbuf *out, struct bch_fs *c,
-			    struct bkey_s_c k)
+			       struct bkey_s_c k)
 {
 	struct bkey_s_c_btree_ptr_v2 bp = bkey_s_c_to_btree_ptr_v2(k);
 
-	pr_buf(out, "seq %llx written %u min_key %s",
+	prt_printf(out, "seq %llx written %u min_key %s",
 	       le64_to_cpu(bp.v->seq),
 	       le16_to_cpu(bp.v->sectors_written),
 	       BTREE_PTR_RANGE_UPDATED(bp.v) ? "R " : "");
 
 	bch2_bpos_to_text(out, bp.v->min_key);
-	pr_buf(out, " ");
+	prt_printf(out, " ");
 	bch2_bkey_ptrs_to_text(out, c, k);
 }
 
@@ -211,8 +224,8 @@ void bch2_btree_ptr_v2_compat(enum btree_id btree_id, unsigned version,
 	compat_bpos(0, btree_id, version, big_endian, write, &bp.v->min_key);
 
 	if (version < bcachefs_metadata_version_inode_btree_change &&
-	    btree_node_type_is_extents(btree_id) &&
-	    bkey_cmp(bp.v->min_key, POS_MIN))
+	    btree_id_is_extents(btree_id) &&
+	    !bkey_eq(bp.v->min_key, POS_MIN))
 		bp.v->min_key = write
 			? bpos_nosnap_predecessor(bp.v->min_key)
 			: bpos_nosnap_successor(bp.v->min_key);
@@ -220,17 +233,6 @@ void bch2_btree_ptr_v2_compat(enum btree_id btree_id, unsigned version,
 
 /* KEY_TYPE_extent: */
 
-const char *bch2_extent_invalid(const struct bch_fs *c, struct bkey_s_c k)
-{
-	return bch2_bkey_ptrs_invalid(c, k);
-}
-
-void bch2_extent_to_text(struct printbuf *out, struct bch_fs *c,
-			 struct bkey_s_c k)
-{
-	bch2_bkey_ptrs_to_text(out, c, k);
-}
-
 bool bch2_extent_merge(struct bch_fs *c, struct bkey_s l, struct bkey_s_c r)
 {
 	struct bkey_ptrs   l_ptrs = bch2_bkey_ptrs(l);
@@ -265,6 +267,7 @@ bool bch2_extent_merge(struct bch_fs *c, struct bkey_s l, struct bkey_s_c r)
 		    rp.ptr.offset + rp.crc.offset ||
 		    lp.ptr.dev			!= rp.ptr.dev ||
 		    lp.ptr.gen			!= rp.ptr.gen ||
+		    lp.ptr.unwritten		!= rp.ptr.unwritten ||
 		    lp.has_ec			!= rp.has_ec)
 			return false;
 
@@ -287,7 +290,7 @@ bool bch2_extent_merge(struct bch_fs *c, struct bkey_s l, struct bkey_s_c r)
 		if (lp.crc.offset + lp.crc.live_size + rp.crc.live_size <=
 		    lp.crc.uncompressed_size) {
 			/* can use left extent's crc entry */
-		} else if (lp.crc.live_size <= rp.crc.offset ) {
+		} else if (lp.crc.live_size <= rp.crc.offset) {
 			/* can use right extent's crc entry */
 		} else {
 			/* check if checksums can be merged: */
@@ -303,10 +306,22 @@ bool bch2_extent_merge(struct bch_fs *c, struct bkey_s l, struct bkey_s_c r)
 
 			if (lp.crc.csum_type &&
 			    lp.crc.uncompressed_size +
-			    rp.crc.uncompressed_size > c->sb.encoded_extent_max)
+			    rp.crc.uncompressed_size > (c->opts.encoded_extent_max >> 9))
 				return false;
+		}
+
+		en_l = extent_entry_next(en_l);
+		en_r = extent_entry_next(en_r);
+	}
+
+	en_l = l_ptrs.start;
+	en_r = r_ptrs.start;
+	while (en_l < l_ptrs.end && en_r < r_ptrs.end) {
+		if (extent_entry_is_crc(en_l)) {
+			struct bch_extent_crc_unpacked crc_l = bch2_extent_crc_unpack(l.k, entry_to_crc(en_l));
+			struct bch_extent_crc_unpacked crc_r = bch2_extent_crc_unpack(r.k, entry_to_crc(en_r));
 
-			if (lp.crc.uncompressed_size + rp.crc.uncompressed_size >
+			if (crc_l.uncompressed_size + crc_r.uncompressed_size >
 			    bch2_crc_field_size_max[extent_entry_type(en_l)])
 				return false;
 		}
@@ -334,7 +349,7 @@ bool bch2_extent_merge(struct bch_fs *c, struct bkey_s l, struct bkey_s_c r)
 			if (crc_l.offset + crc_l.live_size + crc_r.live_size <=
 			    crc_l.uncompressed_size) {
 				/* can use left extent's crc entry */
-			} else if (crc_l.live_size <= crc_r.offset ) {
+			} else if (crc_l.live_size <= crc_r.offset) {
 				/* can use right extent's crc entry */
 				crc_r.offset -= crc_l.live_size;
 				bch2_extent_crc_pack(entry_to_crc(en_l), crc_r,
@@ -363,17 +378,18 @@ bool bch2_extent_merge(struct bch_fs *c, struct bkey_s l, struct bkey_s_c r)
 
 /* KEY_TYPE_reservation: */
 
-const char *bch2_reservation_invalid(const struct bch_fs *c, struct bkey_s_c k)
+int bch2_reservation_invalid(struct bch_fs *c, struct bkey_s_c k,
+			     enum bkey_invalid_flags flags,
+			     struct printbuf *err)
 {
 	struct bkey_s_c_reservation r = bkey_s_c_to_reservation(k);
+	int ret = 0;
 
-	if (bkey_val_bytes(k.k) != sizeof(struct bch_reservation))
-		return "incorrect value size";
-
-	if (!r.v->nr_replicas || r.v->nr_replicas > BCH_REPLICAS_MAX)
-		return "invalid nr_replicas";
-
-	return NULL;
+	bkey_fsck_err_on(!r.v->nr_replicas || r.v->nr_replicas > BCH_REPLICAS_MAX, c, err,
+			 reservation_key_nr_replicas_invalid,
+			 "invalid nr_replicas (%u)", r.v->nr_replicas);
+fsck_err:
+	return ret;
 }
 
 void bch2_reservation_to_text(struct printbuf *out, struct bch_fs *c,
@@ -381,7 +397,7 @@ void bch2_reservation_to_text(struct printbuf *out, struct bch_fs *c,
 {
 	struct bkey_s_c_reservation r = bkey_s_c_to_reservation(k);
 
-	pr_buf(out, "generation %u replicas %u",
+	prt_printf(out, "generation %u replicas %u",
 	       le32_to_cpu(r.v->generation),
 	       r.v->nr_replicas);
 }
@@ -480,7 +496,7 @@ restart_narrow_pointers:
 
 	bkey_for_each_ptr_decode(&k->k, ptrs, p, i)
 		if (can_narrow_crc(p.crc, n)) {
-			bch2_bkey_drop_ptr(bkey_i_to_s(k), &i->ptr);
+			bch2_bkey_drop_ptr_noerror(bkey_i_to_s(k), &i->ptr);
 			p.ptr.offset += p.crc.offset;
 			p.crc = n;
 			bch2_extent_ptr_decoded_append(k, &p);
@@ -506,13 +522,13 @@ static void bch2_extent_crc_pack(union bch_extent_crc *dst,
 	switch (type) {
 	case BCH_EXTENT_ENTRY_crc32:
 		set_common_fields(dst->crc32, src);
-		dst->crc32.csum	 = *((__le32 *) &src.csum.lo);
+		dst->crc32.csum		= (u32 __force) *((__le32 *) &src.csum.lo);
 		break;
 	case BCH_EXTENT_ENTRY_crc64:
 		set_common_fields(dst->crc64, src);
 		dst->crc64.nonce	= src.nonce;
-		dst->crc64.csum_lo	= src.csum.lo;
-		dst->crc64.csum_hi	= *((__le16 *) &src.csum.hi);
+		dst->crc64.csum_lo	= (u64 __force) src.csum.lo;
+		dst->crc64.csum_hi	= (u64 __force) *((__le16 *) &src.csum.hi);
 		break;
 	case BCH_EXTENT_ENTRY_crc128:
 		set_common_fields(dst->crc128, src);
@@ -612,38 +628,6 @@ bool bch2_bkey_is_incompressible(struct bkey_s_c k)
 	return false;
 }
 
-bool bch2_check_range_allocated(struct bch_fs *c, struct bpos pos, u64 size,
-				unsigned nr_replicas, bool compressed)
-{
-	struct btree_trans trans;
-	struct btree_iter iter;
-	struct bpos end = pos;
-	struct bkey_s_c k;
-	bool ret = true;
-	int err;
-
-	end.offset += size;
-
-	bch2_trans_init(&trans, c, 0, 0);
-
-	for_each_btree_key(&trans, iter, BTREE_ID_extents, pos,
-			   BTREE_ITER_SLOTS, k, err) {
-		if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
-			break;
-
-		if (nr_replicas > bch2_bkey_replicas(c, k) ||
-		    (!compressed && bch2_bkey_sectors_compressed(k))) {
-			ret = false;
-			break;
-		}
-	}
-	bch2_trans_iter_exit(&trans, &iter);
-
-	bch2_trans_exit(&trans);
-
-	return ret;
-}
-
 unsigned bch2_bkey_replicas(struct bch_fs *c, struct bkey_s_c k)
 {
 	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
@@ -665,24 +649,31 @@ unsigned bch2_bkey_replicas(struct bch_fs *c, struct bkey_s_c k)
 	return replicas;
 }
 
-static unsigned bch2_extent_ptr_durability(struct bch_fs *c,
-					   struct extent_ptr_decoded p)
+static inline unsigned __extent_ptr_durability(struct bch_dev *ca, struct extent_ptr_decoded *p)
 {
-	unsigned durability = 0;
-	struct bch_dev *ca;
-
-	if (p.ptr.cached)
+	if (p->ptr.cached)
 		return 0;
 
-	ca = bch_dev_bkey_exists(c, p.ptr.dev);
+	return p->has_ec
+		? p->ec.redundancy + 1
+		: ca->mi.durability;
+}
 
-	if (ca->mi.state != BCH_MEMBER_STATE_failed)
-		durability = max_t(unsigned, durability, ca->mi.durability);
+unsigned bch2_extent_ptr_desired_durability(struct bch_fs *c, struct extent_ptr_decoded *p)
+{
+	struct bch_dev *ca = bch_dev_bkey_exists(c, p->ptr.dev);
 
-	if (p.has_ec)
-		durability += p.ec.redundancy;
+	return __extent_ptr_durability(ca, p);
+}
 
-	return durability;
+unsigned bch2_extent_ptr_durability(struct bch_fs *c, struct extent_ptr_decoded *p)
+{
+	struct bch_dev *ca = bch_dev_bkey_exists(c, p->ptr.dev);
+
+	if (ca->mi.state == BCH_MEMBER_STATE_failed)
+		return 0;
+
+	return __extent_ptr_durability(ca, p);
 }
 
 unsigned bch2_bkey_durability(struct bch_fs *c, struct bkey_s_c k)
@@ -693,40 +684,23 @@ unsigned bch2_bkey_durability(struct bch_fs *c, struct bkey_s_c k)
 	unsigned durability = 0;
 
 	bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
-		durability += bch2_extent_ptr_durability(c, p);
+		durability += bch2_extent_ptr_durability(c, &p);
 
 	return durability;
 }
 
-void bch2_bkey_mark_replicas_cached(struct bch_fs *c, struct bkey_s k,
-				    unsigned target,
-				    unsigned nr_desired_replicas)
+static unsigned bch2_bkey_durability_safe(struct bch_fs *c, struct bkey_s_c k)
 {
-	struct bkey_ptrs ptrs = bch2_bkey_ptrs(k);
-	union bch_extent_entry *entry;
+	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
+	const union bch_extent_entry *entry;
 	struct extent_ptr_decoded p;
-	int extra = bch2_bkey_durability(c, k.s_c) - nr_desired_replicas;
-
-	if (target && extra > 0)
-		bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
-			int n = bch2_extent_ptr_durability(c, p);
-
-			if (n && n <= extra &&
-			    !bch2_dev_in_target(c, p.ptr.dev, target)) {
-				entry->ptr.cached = true;
-				extra -= n;
-			}
-		}
+	unsigned durability = 0;
 
-	if (extra > 0)
-		bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
-			int n = bch2_extent_ptr_durability(c, p);
+	bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
+		if (p.ptr.dev < c->sb.nr_devices && c->devs[p.ptr.dev])
+			durability += bch2_extent_ptr_durability(c, &p);
 
-			if (n && n <= extra) {
-				entry->ptr.cached = true;
-				extra -= n;
-			}
-		}
+	return durability;
 }
 
 void bch2_bkey_extent_entry_drop(struct bkey_i *k, union bch_extent_entry *entry)
@@ -738,41 +712,6 @@ void bch2_bkey_extent_entry_drop(struct bkey_i *k, union bch_extent_entry *entry
 	k->k.u64s -= extent_entry_u64s(entry);
 }
 
-void bch2_bkey_append_ptr(struct bkey_i *k,
-			  struct bch_extent_ptr ptr)
-{
-	EBUG_ON(bch2_bkey_has_device(bkey_i_to_s_c(k), ptr.dev));
-
-	switch (k->k.type) {
-	case KEY_TYPE_btree_ptr:
-	case KEY_TYPE_btree_ptr_v2:
-	case KEY_TYPE_extent:
-		EBUG_ON(bkey_val_u64s(&k->k) >= BKEY_EXTENT_VAL_U64s_MAX);
-
-		ptr.type = 1 << BCH_EXTENT_ENTRY_ptr;
-
-		memcpy((void *) &k->v + bkey_val_bytes(&k->k),
-		       &ptr,
-		       sizeof(ptr));
-		k->u64s++;
-		break;
-	default:
-		BUG();
-	}
-}
-
-static inline void __extent_entry_insert(struct bkey_i *k,
-					 union bch_extent_entry *dst,
-					 union bch_extent_entry *new)
-{
-	union bch_extent_entry *end = bkey_val_end(bkey_i_to_s(k));
-
-	memmove_u64s_up_small((u64 *) dst + extent_entry_u64s(new),
-			      dst, (u64 *) end - (u64 *) dst);
-	k->k.u64s += extent_entry_u64s(new);
-	memcpy(dst, new, extent_entry_bytes(new));
-}
-
 void bch2_extent_ptr_decoded_append(struct bkey_i *k,
 				    struct extent_ptr_decoded *p)
 {
@@ -817,41 +756,73 @@ static union bch_extent_entry *extent_entry_prev(struct bkey_ptrs ptrs,
 	return i;
 }
 
-union bch_extent_entry *bch2_bkey_drop_ptr(struct bkey_s k,
-					   struct bch_extent_ptr *ptr)
+/*
+ * Returns pointer to the next entry after the one being dropped:
+ */
+union bch_extent_entry *bch2_bkey_drop_ptr_noerror(struct bkey_s k,
+						   struct bch_extent_ptr *ptr)
 {
 	struct bkey_ptrs ptrs = bch2_bkey_ptrs(k);
-	union bch_extent_entry *dst, *src, *prev;
+	union bch_extent_entry *entry = to_entry(ptr), *next;
+	union bch_extent_entry *ret = entry;
 	bool drop_crc = true;
 
 	EBUG_ON(ptr < &ptrs.start->ptr ||
 		ptr >= &ptrs.end->ptr);
 	EBUG_ON(ptr->type != 1 << BCH_EXTENT_ENTRY_ptr);
 
-	src = extent_entry_next(to_entry(ptr));
-	if (src != ptrs.end &&
-	    !extent_entry_is_crc(src))
-		drop_crc = false;
-
-	dst = to_entry(ptr);
-	while ((prev = extent_entry_prev(ptrs, dst))) {
-		if (extent_entry_is_ptr(prev))
+	for (next = extent_entry_next(entry);
+	     next != ptrs.end;
+	     next = extent_entry_next(next)) {
+		if (extent_entry_is_crc(next)) {
 			break;
-
-		if (extent_entry_is_crc(prev)) {
-			if (drop_crc)
-				dst = prev;
+		} else if (extent_entry_is_ptr(next)) {
+			drop_crc = false;
 			break;
 		}
+	}
+
+	extent_entry_drop(k, entry);
+
+	while ((entry = extent_entry_prev(ptrs, entry))) {
+		if (extent_entry_is_ptr(entry))
+			break;
 
-		dst = prev;
+		if ((extent_entry_is_crc(entry) && drop_crc) ||
+		    extent_entry_is_stripe_ptr(entry)) {
+			ret = (void *) ret - extent_entry_bytes(entry);
+			extent_entry_drop(k, entry);
+		}
 	}
 
-	memmove_u64s_down(dst, src,
-			  (u64 *) ptrs.end - (u64 *) src);
-	k.k->u64s -= (u64 *) src - (u64 *) dst;
+	return ret;
+}
 
-	return dst;
+union bch_extent_entry *bch2_bkey_drop_ptr(struct bkey_s k,
+					   struct bch_extent_ptr *ptr)
+{
+	bool have_dirty = bch2_bkey_dirty_devs(k.s_c).nr;
+	union bch_extent_entry *ret =
+		bch2_bkey_drop_ptr_noerror(k, ptr);
+
+	/*
+	 * If we deleted all the dirty pointers and there's still cached
+	 * pointers, we could set the cached pointers to dirty if they're not
+	 * stale - but to do that correctly we'd need to grab an open_bucket
+	 * reference so that we don't race with bucket reuse:
+	 */
+	if (have_dirty &&
+	    !bch2_bkey_dirty_devs(k.s_c).nr) {
+		k.k->type = KEY_TYPE_error;
+		set_bkey_val_u64s(k.k, 0);
+		ret = NULL;
+	} else if (!bch2_bkey_nr_ptrs(k.s_c)) {
+		k.k->type = KEY_TYPE_deleted;
+		set_bkey_val_u64s(k.k, 0);
+		ret = NULL;
+	}
+
+	return ret;
 }
 
 void bch2_bkey_drop_device(struct bkey_s k, unsigned dev)
@@ -861,11 +832,17 @@ void bch2_bkey_drop_device(struct bkey_s k, unsigned dev)
 	bch2_bkey_drop_ptrs(k, ptr, ptr->dev == dev);
 }
 
-const struct bch_extent_ptr *
-bch2_bkey_has_device(struct bkey_s_c k, unsigned dev)
+void bch2_bkey_drop_device_noerror(struct bkey_s k, unsigned dev)
+{
+	struct bch_extent_ptr *ptr = bch2_bkey_has_device(k, dev);
+
+	if (ptr)
+		bch2_bkey_drop_ptr_noerror(k, ptr);
+}
+
+const struct bch_extent_ptr *bch2_bkey_has_device_c(struct bkey_s_c k, unsigned dev)
 {
 	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
-	const struct bch_extent_ptr *ptr;
 
 	bkey_for_each_ptr(ptrs, ptr)
 		if (ptr->dev == dev)
@@ -877,7 +854,6 @@ bch2_bkey_has_device(struct bkey_s_c k, unsigned dev)
 bool bch2_bkey_has_target(struct bch_fs *c, struct bkey_s_c k, unsigned target)
 {
 	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
-	const struct bch_extent_ptr *ptr;
 
 	bkey_for_each_ptr(ptrs, ptr)
 		if (bch2_dev_in_target(c, ptr->dev, target) &&
@@ -905,6 +881,78 @@ bool bch2_bkey_matches_ptr(struct bch_fs *c, struct bkey_s_c k,
 	return false;
 }
 
+/*
+ * Returns true if two extents refer to the same data:
+ */
+bool bch2_extents_match(struct bkey_s_c k1, struct bkey_s_c k2)
+{
+	if (k1.k->type != k2.k->type)
+		return false;
+
+	if (bkey_extent_is_direct_data(k1.k)) {
+		struct bkey_ptrs_c ptrs1 = bch2_bkey_ptrs_c(k1);
+		struct bkey_ptrs_c ptrs2 = bch2_bkey_ptrs_c(k2);
+		const union bch_extent_entry *entry1, *entry2;
+		struct extent_ptr_decoded p1, p2;
+
+		if (bkey_extent_is_unwritten(k1) != bkey_extent_is_unwritten(k2))
+			return false;
+
+		bkey_for_each_ptr_decode(k1.k, ptrs1, p1, entry1)
+			bkey_for_each_ptr_decode(k2.k, ptrs2, p2, entry2)
+				if (p1.ptr.dev		== p2.ptr.dev &&
+				    p1.ptr.gen		== p2.ptr.gen &&
+				    (s64) p1.ptr.offset + p1.crc.offset - bkey_start_offset(k1.k) ==
+				    (s64) p2.ptr.offset + p2.crc.offset - bkey_start_offset(k2.k))
+					return true;
+
+		return false;
+	} else {
+		/* KEY_TYPE_deleted, etc. */
+		return true;
+	}
+}
+
+struct bch_extent_ptr *
+bch2_extent_has_ptr(struct bkey_s_c k1, struct extent_ptr_decoded p1, struct bkey_s k2)
+{
+	struct bkey_ptrs ptrs2 = bch2_bkey_ptrs(k2);
+	union bch_extent_entry *entry2;
+	struct extent_ptr_decoded p2;
+
+	bkey_for_each_ptr_decode(k2.k, ptrs2, p2, entry2)
+		if (p1.ptr.dev		== p2.ptr.dev &&
+		    p1.ptr.gen		== p2.ptr.gen &&
+		    (s64) p1.ptr.offset + p1.crc.offset - bkey_start_offset(k1.k) ==
+		    (s64) p2.ptr.offset + p2.crc.offset - bkey_start_offset(k2.k))
+			return &entry2->ptr;
+
+	return NULL;
+}
+
+void bch2_extent_ptr_set_cached(struct bkey_s k, struct bch_extent_ptr *ptr)
+{
+	struct bkey_ptrs ptrs = bch2_bkey_ptrs(k);
+	union bch_extent_entry *entry;
+	union bch_extent_entry *ec = NULL;
+
+	bkey_extent_entry_for_each(ptrs, entry) {
+		if (&entry->ptr == ptr) {
+			ptr->cached = true;
+			if (ec)
+				extent_entry_drop(k, ec);
+			return;
+		}
+
+		if (extent_entry_is_stripe_ptr(entry))
+			ec = entry;
+		else if (extent_entry_is_ptr(entry))
+			ec = NULL;
+	}
+
+	BUG();
+}
+
 /*
  * bch_extent_normalize - clean up an extent, dropping stale pointers etc.
  *
@@ -921,10 +969,6 @@ bool bch2_extent_normalize(struct bch_fs *c, struct bkey_s k)
 		ptr->cached &&
 		ptr_stale(bch_dev_bkey_exists(c, ptr->dev), ptr));
 
-	/* will only happen if all pointers were cached: */
-	if (!bch2_bkey_nr_ptrs(k.s_c))
-		k.k->type = KEY_TYPE_deleted;
-
 	return bkey_deleted(k.k);
 }
 
@@ -933,49 +977,76 @@ void bch2_bkey_ptrs_to_text(struct printbuf *out, struct bch_fs *c,
 {
 	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 	const union bch_extent_entry *entry;
-	struct bch_extent_crc_unpacked crc;
-	const struct bch_extent_ptr *ptr;
-	const struct bch_extent_stripe_ptr *ec;
-	struct bch_dev *ca;
 	bool first = true;
 
+	if (c)
+		prt_printf(out, "durability: %u ", bch2_bkey_durability_safe(c, k));
+
 	bkey_extent_entry_for_each(ptrs, entry) {
 		if (!first)
-			pr_buf(out, " ");
+			prt_printf(out, " ");
 
 		switch (__extent_entry_type(entry)) {
-		case BCH_EXTENT_ENTRY_ptr:
-			ptr = entry_to_ptr(entry);
-			ca = ptr->dev < c->sb.nr_devices && c->devs[ptr->dev]
+		case BCH_EXTENT_ENTRY_ptr: {
+			const struct bch_extent_ptr *ptr = entry_to_ptr(entry);
+			struct bch_dev *ca = c && ptr->dev < c->sb.nr_devices && c->devs[ptr->dev]
 				? bch_dev_bkey_exists(c, ptr->dev)
 				: NULL;
 
-			pr_buf(out, "ptr: %u:%llu gen %u%s%s", ptr->dev,
-			       (u64) ptr->offset, ptr->gen,
-			       ptr->cached ? " cached" : "",
-			       ca && ptr_stale(ca, ptr)
-			       ? " stale" : "");
+			if (!ca) {
+				prt_printf(out, "ptr: %u:%llu gen %u%s", ptr->dev,
+				       (u64) ptr->offset, ptr->gen,
+				       ptr->cached ? " cached" : "");
+			} else {
+				u32 offset;
+				u64 b = sector_to_bucket_and_offset(ca, ptr->offset, &offset);
+
+				prt_printf(out, "ptr: %u:%llu:%u gen %u",
+					   ptr->dev, b, offset, ptr->gen);
+				if (ptr->cached)
+					prt_str(out, " cached");
+				if (ptr->unwritten)
+					prt_str(out, " unwritten");
+				if (ca && ptr_stale(ca, ptr))
+					prt_printf(out, " stale");
+			}
 			break;
+		}
 		case BCH_EXTENT_ENTRY_crc32:
 		case BCH_EXTENT_ENTRY_crc64:
-		case BCH_EXTENT_ENTRY_crc128:
-			crc = bch2_extent_crc_unpack(k.k, entry_to_crc(entry));
+		case BCH_EXTENT_ENTRY_crc128: {
+			struct bch_extent_crc_unpacked crc =
+				bch2_extent_crc_unpack(k.k, entry_to_crc(entry));
 
-			pr_buf(out, "crc: c_size %u size %u offset %u nonce %u csum %u compress %u",
+			prt_printf(out, "crc: c_size %u size %u offset %u nonce %u csum %s compress %s",
 			       crc.compressed_size,
 			       crc.uncompressed_size,
 			       crc.offset, crc.nonce,
-			       crc.csum_type,
-			       crc.compression_type);
+			       bch2_csum_types[crc.csum_type],
+			       bch2_compression_types[crc.compression_type]);
 			break;
-		case BCH_EXTENT_ENTRY_stripe_ptr:
-			ec = &entry->stripe_ptr;
+		}
+		case BCH_EXTENT_ENTRY_stripe_ptr: {
+			const struct bch_extent_stripe_ptr *ec = &entry->stripe_ptr;
 
-			pr_buf(out, "ec: idx %llu block %u",
+			prt_printf(out, "ec: idx %llu block %u",
 			       (u64) ec->idx, ec->block);
 			break;
+		}
+		case BCH_EXTENT_ENTRY_rebalance: {
+			const struct bch_extent_rebalance *r = &entry->rebalance;
+
+			prt_str(out, "rebalance: target ");
+			if (c)
+				bch2_target_to_text(out, c, r->target);
+			else
+				prt_printf(out, "%u", r->target);
+			prt_str(out, " compression ");
+			bch2_compression_opt_to_text(out, r->compression);
+			break;
+		}
 		default:
-			pr_buf(out, "(invalid extent entry %.16llx)", *((u64 *) entry));
+			prt_printf(out, "(invalid extent entry %.16llx)", *((u64 *) entry));
 			return;
 		}
 
@@ -983,106 +1054,178 @@ void bch2_bkey_ptrs_to_text(struct printbuf *out, struct bch_fs *c,
 	}
 }
 
-static const char *extent_ptr_invalid(const struct bch_fs *c,
-				      struct bkey_s_c k,
-				      const struct bch_extent_ptr *ptr,
-				      unsigned size_ondisk,
-				      bool metadata)
+static int extent_ptr_invalid(struct bch_fs *c,
+			      struct bkey_s_c k,
+			      enum bkey_invalid_flags flags,
+			      const struct bch_extent_ptr *ptr,
+			      unsigned size_ondisk,
+			      bool metadata,
+			      struct printbuf *err)
 {
 	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
-	const struct bch_extent_ptr *ptr2;
+	u64 bucket;
+	u32 bucket_offset;
 	struct bch_dev *ca;
+	int ret = 0;
 
-	if (!bch2_dev_exists2(c, ptr->dev))
-		return "pointer to invalid device";
+	if (!bch2_dev_exists2(c, ptr->dev)) {
+		/*
+		 * If we're in the write path this key might have already been
+		 * overwritten, and we could be seeing a device that doesn't
+		 * exist anymore due to racing with device removal:
+		 */
+		if (flags & BKEY_INVALID_WRITE)
+			return 0;
 
-	ca = bch_dev_bkey_exists(c, ptr->dev);
-	if (!ca)
-		return "pointer to invalid device";
+		bkey_fsck_err(c, err, ptr_to_invalid_device,
+			   "pointer to invalid device (%u)", ptr->dev);
+	}
 
+	ca = bch_dev_bkey_exists(c, ptr->dev);
 	bkey_for_each_ptr(ptrs, ptr2)
-		if (ptr != ptr2 && ptr->dev == ptr2->dev)
-			return "multiple pointers to same device";
-
-	if (ptr->offset + size_ondisk > bucket_to_sector(ca, ca->mi.nbuckets))
-		return "offset past end of device";
-
-	if (ptr->offset < bucket_to_sector(ca, ca->mi.first_bucket))
-		return "offset before first bucket";
-
-	if (bucket_remainder(ca, ptr->offset) +
-	    size_ondisk > ca->mi.bucket_size)
-		return "spans multiple buckets";
-
-	return NULL;
+		bkey_fsck_err_on(ptr != ptr2 && ptr->dev == ptr2->dev, c, err,
+				 ptr_to_duplicate_device,
+				 "multiple pointers to same device (%u)", ptr->dev);
+
+	bucket = sector_to_bucket_and_offset(ca, ptr->offset, &bucket_offset);
+
+	bkey_fsck_err_on(bucket >= ca->mi.nbuckets, c, err,
+			 ptr_after_last_bucket,
+			 "pointer past last bucket (%llu > %llu)", bucket, ca->mi.nbuckets);
+	bkey_fsck_err_on(ptr->offset < bucket_to_sector(ca, ca->mi.first_bucket), c, err,
+			 ptr_before_first_bucket,
+			 "pointer before first bucket (%llu < %u)", bucket, ca->mi.first_bucket);
+	bkey_fsck_err_on(bucket_offset + size_ondisk > ca->mi.bucket_size, c, err,
+			 ptr_spans_multiple_buckets,
+			 "pointer spans multiple buckets (%u + %u > %u)",
+		       bucket_offset, size_ondisk, ca->mi.bucket_size);
+fsck_err:
+	return ret;
 }
 
-const char *bch2_bkey_ptrs_invalid(const struct bch_fs *c, struct bkey_s_c k)
+int bch2_bkey_ptrs_invalid(struct bch_fs *c, struct bkey_s_c k,
+			   enum bkey_invalid_flags flags,
+			   struct printbuf *err)
 {
 	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
-	struct bch_devs_list devs;
 	const union bch_extent_entry *entry;
 	struct bch_extent_crc_unpacked crc;
 	unsigned size_ondisk = k.k->size;
-	const char *reason;
 	unsigned nonce = UINT_MAX;
-	unsigned i;
+	unsigned nr_ptrs = 0;
+	bool have_written = false, have_unwritten = false, have_ec = false, crc_since_last_ptr = false;
+	int ret = 0;
 
-	if (k.k->type == KEY_TYPE_btree_ptr ||
-	    k.k->type == KEY_TYPE_btree_ptr_v2)
-		size_ondisk = c->opts.btree_node_size;
+	if (bkey_is_btree_ptr(k.k))
+		size_ondisk = btree_sectors(c);
 
 	bkey_extent_entry_for_each(ptrs, entry) {
-		if (__extent_entry_type(entry) >= BCH_EXTENT_ENTRY_MAX)
-			return "invalid extent entry type";
+		bkey_fsck_err_on(__extent_entry_type(entry) >= BCH_EXTENT_ENTRY_MAX, c, err,
+			extent_ptrs_invalid_entry,
+			"invalid extent entry type (got %u, max %u)",
+			__extent_entry_type(entry), BCH_EXTENT_ENTRY_MAX);
 
-		if (k.k->type == KEY_TYPE_btree_ptr &&
-		    !extent_entry_is_ptr(entry))
-			return "has non ptr field";
+		bkey_fsck_err_on(bkey_is_btree_ptr(k.k) &&
+				 !extent_entry_is_ptr(entry), c, err,
+				 btree_ptr_has_non_ptr,
+				 "has non ptr field");
 
 		switch (extent_entry_type(entry)) {
 		case BCH_EXTENT_ENTRY_ptr:
-			reason = extent_ptr_invalid(c, k, &entry->ptr,
-						    size_ondisk, false);
-			if (reason)
-				return reason;
+			ret = extent_ptr_invalid(c, k, flags, &entry->ptr,
+						 size_ondisk, false, err);
+			if (ret)
+				return ret;
+
+			bkey_fsck_err_on(entry->ptr.cached && have_ec, c, err,
+					 ptr_cached_and_erasure_coded,
+					 "cached, erasure coded ptr");
+
+			if (!entry->ptr.unwritten)
+				have_written = true;
+			else
+				have_unwritten = true;
+
+			have_ec = false;
+			crc_since_last_ptr = false;
+			nr_ptrs++;
 			break;
 		case BCH_EXTENT_ENTRY_crc32:
 		case BCH_EXTENT_ENTRY_crc64:
 		case BCH_EXTENT_ENTRY_crc128:
 			crc = bch2_extent_crc_unpack(k.k, entry_to_crc(entry));
 
-			if (crc.offset + crc.live_size >
-			    crc.uncompressed_size)
-				return "checksum offset + key size > uncompressed size";
-
-			size_ondisk = crc.compressed_size;
-
-			if (!bch2_checksum_type_valid(c, crc.csum_type))
-				return "invalid checksum type";
-
-			if (crc.compression_type >= BCH_COMPRESSION_TYPE_NR)
-				return "invalid compression type";
+			bkey_fsck_err_on(crc.offset + crc.live_size > crc.uncompressed_size, c, err,
+					 ptr_crc_uncompressed_size_too_small,
+					 "checksum offset + key size > uncompressed size");
+			bkey_fsck_err_on(!bch2_checksum_type_valid(c, crc.csum_type), c, err,
+					 ptr_crc_csum_type_unknown,
+					 "invalid checksum type");
+			bkey_fsck_err_on(crc.compression_type >= BCH_COMPRESSION_TYPE_NR, c, err,
+					 ptr_crc_compression_type_unknown,
+					 "invalid compression type");
 
 			if (bch2_csum_type_is_encryption(crc.csum_type)) {
 				if (nonce == UINT_MAX)
 					nonce = crc.offset + crc.nonce;
 				else if (nonce != crc.offset + crc.nonce)
-					return "incorrect nonce";
+					bkey_fsck_err(c, err, ptr_crc_nonce_mismatch,
+						      "incorrect nonce");
 			}
+
+			bkey_fsck_err_on(crc_since_last_ptr, c, err,
+					 ptr_crc_redundant,
+					 "redundant crc entry");
+			crc_since_last_ptr = true;
+
+			bkey_fsck_err_on(crc_is_encoded(crc) &&
+					 (crc.uncompressed_size > c->opts.encoded_extent_max >> 9) &&
+					 (flags & (BKEY_INVALID_WRITE|BKEY_INVALID_COMMIT)), c, err,
+					 ptr_crc_uncompressed_size_too_big,
+					 "too large encoded extent");
+
+			size_ondisk = crc.compressed_size;
 			break;
 		case BCH_EXTENT_ENTRY_stripe_ptr:
+			bkey_fsck_err_on(have_ec, c, err,
+					 ptr_stripe_redundant,
+					 "redundant stripe entry");
+			have_ec = true;
 			break;
+		case BCH_EXTENT_ENTRY_rebalance: {
+			const struct bch_extent_rebalance *r = &entry->rebalance;
+
+			if (!bch2_compression_opt_valid(r->compression)) {
+				struct bch_compression_opt opt = __bch2_compression_decode(r->compression);
+				prt_printf(err, "invalid compression opt %u:%u",
+					   opt.type, opt.level);
+				return -BCH_ERR_invalid_bkey;
+			}
+			break;
+		}
 		}
 	}
 
-	devs = bch2_bkey_devs(k);
-	bubble_sort(devs.devs, devs.nr, u8_cmp);
-	for (i = 0; i + 1 < devs.nr; i++)
-		if (devs.devs[i] == devs.devs[i + 1])
-			return "multiple ptrs to same device";
-
-	return NULL;
+	bkey_fsck_err_on(!nr_ptrs, c, err,
+			 extent_ptrs_no_ptrs,
+			 "no ptrs");
+	bkey_fsck_err_on(nr_ptrs > BCH_BKEY_PTRS_MAX, c, err,
+			 extent_ptrs_too_many_ptrs,
+			 "too many ptrs: %u > %u", nr_ptrs, BCH_BKEY_PTRS_MAX);
+	bkey_fsck_err_on(have_written && have_unwritten, c, err,
+			 extent_ptrs_written_and_unwritten,
+			 "extent with unwritten and written ptrs");
+	bkey_fsck_err_on(k.k->type != KEY_TYPE_extent && have_unwritten, c, err,
+			 extent_ptrs_unwritten,
+			 "has unwritten ptrs");
+	bkey_fsck_err_on(crc_since_last_ptr, c, err,
+			 extent_ptrs_redundant_crc,
+			 "redundant crc entry");
+	bkey_fsck_err_on(have_ec, c, err,
+			 extent_ptrs_redundant_stripe,
+			 "redundant stripe entry");
+fsck_err:
+	return ret;
 }
 
 void bch2_ptr_swab(struct bkey_s k)
@@ -1117,8 +1260,129 @@ void bch2_ptr_swab(struct bkey_s k)
 			break;
 		case BCH_EXTENT_ENTRY_stripe_ptr:
 			break;
+		case BCH_EXTENT_ENTRY_rebalance:
+			break;
+		}
+	}
+}
+
+const struct bch_extent_rebalance *bch2_bkey_rebalance_opts(struct bkey_s_c k)
+{
+	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
+	const union bch_extent_entry *entry;
+
+	bkey_extent_entry_for_each(ptrs, entry)
+		if (__extent_entry_type(entry) == BCH_EXTENT_ENTRY_rebalance)
+			return &entry->rebalance;
+
+	return NULL;
+}
+
+unsigned bch2_bkey_ptrs_need_rebalance(struct bch_fs *c, struct bkey_s_c k,
+				       unsigned target, unsigned compression)
+{
+	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
+	unsigned rewrite_ptrs = 0;
+
+	if (compression) {
+		unsigned compression_type = bch2_compression_opt_to_type(compression);
+		const union bch_extent_entry *entry;
+		struct extent_ptr_decoded p;
+		unsigned i = 0;
+
+		bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
+			if (p.crc.compression_type == BCH_COMPRESSION_TYPE_incompressible ||
+			    p.ptr.unwritten) {
+				rewrite_ptrs = 0;
+				goto incompressible;
+			}
+
+			if (!p.ptr.cached && p.crc.compression_type != compression_type)
+				rewrite_ptrs |= 1U << i;
+			i++;
+		}
+	}
+incompressible:
+	if (target && bch2_target_accepts_data(c, BCH_DATA_user, target)) {
+		unsigned i = 0;
+
+		bkey_for_each_ptr(ptrs, ptr) {
+			if (!ptr->cached && !bch2_dev_in_target(c, ptr->dev, target))
+				rewrite_ptrs |= 1U << i;
+			i++;
 		}
 	}
+
+	return rewrite_ptrs;
+}
+
+bool bch2_bkey_needs_rebalance(struct bch_fs *c, struct bkey_s_c k)
+{
+	const struct bch_extent_rebalance *r = bch2_bkey_rebalance_opts(k);
+
+	/*
+	 * If it's an indirect extent, we don't delete the rebalance entry when
+	 * done so that we know what options were applied - check if it still
+	 * needs work done:
+	 */
+	if (r &&
+	    k.k->type == KEY_TYPE_reflink_v &&
+	    !bch2_bkey_ptrs_need_rebalance(c, k, r->target, r->compression))
+		r = NULL;
+
+	return r != NULL;
+}
+
+int bch2_bkey_set_needs_rebalance(struct bch_fs *c, struct bkey_i *_k,
+				  unsigned target, unsigned compression)
+{
+	struct bkey_s k = bkey_i_to_s(_k);
+	struct bch_extent_rebalance *r;
+	bool needs_rebalance;
+
+	if (!bkey_extent_is_direct_data(k.k))
+		return 0;
+
+	/* get existing rebalance entry: */
+	r = (struct bch_extent_rebalance *) bch2_bkey_rebalance_opts(k.s_c);
+	if (r) {
+		if (k.k->type == KEY_TYPE_reflink_v) {
+			/*
+			 * indirect extents: existing options take precedence,
+			 * so that we don't move extents back and forth if
+			 * they're referenced by different inodes with different
+			 * options:
+			 */
+			if (r->target)
+				target = r->target;
+			if (r->compression)
+				compression = r->compression;
+		}
+
+		r->target	= target;
+		r->compression	= compression;
+	}
+
+	needs_rebalance = bch2_bkey_ptrs_need_rebalance(c, k.s_c, target, compression);
+
+	if (needs_rebalance && !r) {
+		union bch_extent_entry *new = bkey_val_end(k);
+
+		new->rebalance.type		= 1U << BCH_EXTENT_ENTRY_rebalance;
+		new->rebalance.compression	= compression;
+		new->rebalance.target		= target;
+		new->rebalance.unused		= 0;
+		k.k->u64s += extent_entry_u64s(new);
+	} else if (!needs_rebalance && r && k.k->type != KEY_TYPE_reflink_v) {
+		/*
+		 * For indirect extents, don't delete the rebalance entry when
+		 * we're finished so that we know we specifically moved it or
+		 * compressed it to its current location/compression type
+		 */
+		extent_entry_drop(k, (union bch_extent_entry *) r);
+	}
+
+	return 0;
 }
 
 /* Generic extent code: */
@@ -1129,10 +1393,10 @@ int bch2_cut_front_s(struct bpos where, struct bkey_s k)
 	int val_u64s_delta;
 	u64 sub;
 
-	if (bkey_cmp(where, bkey_start_pos(k.k)) <= 0)
+	if (bkey_le(where, bkey_start_pos(k.k)))
 		return 0;
 
-	EBUG_ON(bkey_cmp(where, k.k->p) > 0);
+	EBUG_ON(bkey_gt(where, k.k->p));
 
 	sub = where.offset - bkey_start_offset(k.k);
 
@@ -1167,6 +1431,8 @@ int bch2_cut_front_s(struct bpos where, struct bkey_s k)
 				break;
 			case BCH_EXTENT_ENTRY_stripe_ptr:
 				break;
+			case BCH_EXTENT_ENTRY_rebalance:
+				break;
 			}
 
 			if (extent_entry_is_crc(entry))
@@ -1209,10 +1475,10 @@ int bch2_cut_back_s(struct bpos where, struct bkey_s k)
 	int val_u64s_delta;
 	u64 len = 0;
 
-	if (bkey_cmp(where, k.k->p) >= 0)
+	if (bkey_ge(where, k.k->p))
 		return 0;
 
-	EBUG_ON(bkey_cmp(where, bkey_start_pos(k.k)) < 0);
+	EBUG_ON(bkey_lt(where, bkey_start_pos(k.k)));
 
 	len = where.offset - bkey_start_offset(k.k);