X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libbcachefs%2Falloc_foreground.c;h=a961df74d421707468ed0868932d25a59fc218fb;hb=e3e7f67b3ed89f5b3158142c29e66bb98f868ce2;hp=a32aaf51bb3e36dac0d1c13efbd3f19e3e3df58c;hpb=934a84dfaf719af82dadbbe0e2480baff03c905b;p=bcachefs-tools-debian

diff --git a/libbcachefs/alloc_foreground.c b/libbcachefs/alloc_foreground.c
index a32aaf5..a961df7 100644
--- a/libbcachefs/alloc_foreground.c
+++ b/libbcachefs/alloc_foreground.c
@@ -25,18 +25,28 @@
 #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"
+#include "trace.h"
 
 #include <linux/math64.h>
 #include <linux/rculist.h>
 #include <linux/rcupdate.h>
-#include <trace/events/bcachefs.h>
 
-const char * const bch2_alloc_reserves[] = {
+static void bch2_trans_mutex_lock_norelock(struct btree_trans *trans,
+					   struct mutex *lock)
+{
+	if (!mutex_trylock(lock)) {
+		bch2_trans_unlock(trans);
+		mutex_lock(lock);
+	}
+}
+
+const char * const bch2_watermarks[] = {
 #define x(t) #t,
-	BCH_ALLOC_RESERVES()
+	BCH_WATERMARKS()
 #undef x
 	NULL
 };
@@ -57,6 +67,17 @@ const char * const bch2_alloc_reserves[] = {
  * reference _after_ doing the index update that makes its allocation reachable.
  */
 
+void bch2_reset_alloc_cursors(struct bch_fs *c)
+{
+	struct bch_dev *ca;
+	unsigned i;
+
+	rcu_read_lock();
+	for_each_member_device_rcu(ca, c, i, NULL)
+		ca->alloc_cursor = 0;
+	rcu_read_unlock();
+}
+
 static void bch2_open_bucket_hash_add(struct bch_fs *c, struct open_bucket *ob)
 {
 	open_bucket_idx_t idx = ob - c->open_buckets;
@@ -85,7 +106,7 @@ void __bch2_open_bucket_put(struct bch_fs *c, struct open_bucket *ob)
 	struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
 
 	if (ob->ec) {
-		bch2_ec_bucket_written(c, ob);
+		ec_stripe_new_put(c, ob->ec, STRIPE_REF_io);
 		return;
 	}
 
@@ -138,30 +159,19 @@ static struct open_bucket *bch2_open_bucket_alloc(struct bch_fs *c)
 	return ob;
 }
 
-static void open_bucket_free_unused(struct bch_fs *c,
-				    struct write_point *wp,
-				    struct open_bucket *ob)
+static void open_bucket_free_unused(struct bch_fs *c, struct open_bucket *ob)
 {
-	struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
-	bool may_realloc = wp->data_type == BCH_DATA_user;
-
-	BUG_ON(ca->open_buckets_partial_nr >
-	       ARRAY_SIZE(ca->open_buckets_partial));
-
-	if (ca->open_buckets_partial_nr <
-	    ARRAY_SIZE(ca->open_buckets_partial) &&
-	    may_realloc) {
-		spin_lock(&c->freelist_lock);
-		ob->on_partial_list = true;
-		ca->open_buckets_partial[ca->open_buckets_partial_nr++] =
-			ob - c->open_buckets;
-		spin_unlock(&c->freelist_lock);
+	BUG_ON(c->open_buckets_partial_nr >=
+	       ARRAY_SIZE(c->open_buckets_partial));
 
-		closure_wake_up(&c->open_buckets_wait);
-		closure_wake_up(&c->freelist_wait);
-	} else {
-		bch2_open_bucket_put(c, ob);
-	}
+	spin_lock(&c->freelist_lock);
+	ob->on_partial_list = true;
+	c->open_buckets_partial[c->open_buckets_partial_nr++] =
+		ob - c->open_buckets;
+	spin_unlock(&c->freelist_lock);
+
+	closure_wake_up(&c->open_buckets_wait);
+	closure_wake_up(&c->freelist_wait);
 }
 
 /* _only_ for allocating the journal on a new device: */
@@ -178,14 +188,16 @@ long bch2_bucket_alloc_new_fs(struct bch_dev *ca)
 	return -1;
 }
 
-static inline unsigned open_buckets_reserved(enum alloc_reserve reserve)
+static inline unsigned open_buckets_reserved(enum bch_watermark watermark)
 {
-	switch (reserve) {
-	case RESERVE_btree:
-	case RESERVE_btree_movinggc:
+	switch (watermark) {
+	case BCH_WATERMARK_reclaim:
 		return 0;
-	case RESERVE_movinggc:
+	case BCH_WATERMARK_btree:
+	case BCH_WATERMARK_btree_copygc:
 		return OPEN_BUCKETS_COUNT / 4;
+	case BCH_WATERMARK_copygc:
+		return OPEN_BUCKETS_COUNT / 3;
 	default:
 		return OPEN_BUCKETS_COUNT / 2;
 	}
@@ -193,8 +205,8 @@ static inline unsigned open_buckets_reserved(enum alloc_reserve reserve)
 
 static struct open_bucket *__try_alloc_bucket(struct bch_fs *c, struct bch_dev *ca,
 					      u64 bucket,
-					      enum alloc_reserve reserve,
-					      struct bch_alloc_v4 *a,
+					      enum bch_watermark watermark,
+					      const struct bch_alloc_v4 *a,
 					      struct bucket_alloc_state *s,
 					      struct closure *cl)
 {
@@ -223,7 +235,7 @@ static struct open_bucket *__try_alloc_bucket(struct bch_fs *c, struct bch_dev *
 
 	spin_lock(&c->freelist_lock);
 
-	if (unlikely(c->open_buckets_nr_free <= open_buckets_reserved(reserve))) {
+	if (unlikely(c->open_buckets_nr_free <= open_buckets_reserved(watermark))) {
 		if (cl)
 			closure_wait(&c->open_buckets_wait, cl);
 
@@ -247,7 +259,6 @@ static struct open_bucket *__try_alloc_bucket(struct bch_fs *c, struct bch_dev *
 
 	ob->valid	= true;
 	ob->sectors_free = ca->mi.bucket_size;
-	ob->alloc_reserve = reserve;
 	ob->dev		= ca->dev_idx;
 	ob->gen		= a->gen;
 	ob->bucket	= bucket;
@@ -271,12 +282,11 @@ static struct open_bucket *__try_alloc_bucket(struct bch_fs *c, struct bch_dev *
 	}
 
 	spin_unlock(&c->freelist_lock);
-
 	return ob;
 }
 
 static struct open_bucket *try_alloc_bucket(struct btree_trans *trans, struct bch_dev *ca,
-					    enum alloc_reserve reserve, u64 free_entry,
+					    enum bch_watermark watermark, u64 free_entry,
 					    struct bucket_alloc_state *s,
 					    struct bkey_s_c freespace_k,
 					    struct closure *cl)
@@ -285,7 +295,8 @@ static struct open_bucket *try_alloc_bucket(struct btree_trans *trans, struct bc
 	struct btree_iter iter = { NULL };
 	struct bkey_s_c k;
 	struct open_bucket *ob;
-	struct bch_alloc_v4 a;
+	struct bch_alloc_v4 a_convert;
+	const struct bch_alloc_v4 *a;
 	u64 b = free_entry & ~(~0ULL << 56);
 	unsigned genbits = free_entry >> 56;
 	struct printbuf buf = PRINTBUF;
@@ -301,32 +312,38 @@ static struct open_bucket *try_alloc_bucket(struct btree_trans *trans, struct bc
 		goto err;
 	}
 
-	bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc, POS(ca->dev_idx, b), BTREE_ITER_CACHED);
-	k = bch2_btree_iter_peek_slot(&iter);
+	k = bch2_bkey_get_iter(trans, &iter,
+			       BTREE_ID_alloc, POS(ca->dev_idx, b),
+			       BTREE_ITER_CACHED);
 	ret = bkey_err(k);
 	if (ret) {
 		ob = ERR_PTR(ret);
 		goto err;
 	}
 
-	bch2_alloc_to_v4(k, &a);
+	a = bch2_alloc_to_v4(k, &a_convert);
 
-	if (genbits != (alloc_freespace_genbits(a) >> 56)) {
-		prt_printf(&buf, "bucket in freespace btree with wrong genbits (got %u should be %llu)\n"
-		       "  freespace key ",
-		       genbits, alloc_freespace_genbits(a) >> 56);
+	if (a->data_type != BCH_DATA_free) {
+		if (c->curr_recovery_pass <= BCH_RECOVERY_PASS_check_alloc_info) {
+			ob = NULL;
+			goto err;
+		}
+
+		prt_printf(&buf, "non free bucket in freespace btree\n"
+		       "  freespace key ");
 		bch2_bkey_val_to_text(&buf, c, freespace_k);
 		prt_printf(&buf, "\n  ");
 		bch2_bkey_val_to_text(&buf, c, k);
 		bch2_trans_inconsistent(trans, "%s", buf.buf);
 		ob = ERR_PTR(-EIO);
 		goto err;
-
 	}
 
-	if (a.data_type != BCH_DATA_free) {
-		prt_printf(&buf, "non free bucket in freespace btree\n"
-		       "  freespace key ");
+	if (genbits != (alloc_freespace_genbits(*a) >> 56) &&
+	    c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_info) {
+		prt_printf(&buf, "bucket in freespace btree with wrong genbits (got %u should be %llu)\n"
+		       "  freespace key ",
+		       genbits, alloc_freespace_genbits(*a) >> 56);
 		bch2_bkey_val_to_text(&buf, c, freespace_k);
 		prt_printf(&buf, "\n  ");
 		bch2_bkey_val_to_text(&buf, c, k);
@@ -335,19 +352,19 @@ static struct open_bucket *try_alloc_bucket(struct btree_trans *trans, struct bc
 		goto err;
 	}
 
-	if (!test_bit(BCH_FS_CHECK_BACKPOINTERS_DONE, &c->flags)) {
+	if (c->curr_recovery_pass <= BCH_RECOVERY_PASS_check_extents_to_backpointers) {
 		struct bch_backpointer bp;
-		u64 bp_offset = 0;
+		struct bpos bp_pos = POS_MIN;
 
 		ret = bch2_get_next_backpointer(trans, POS(ca->dev_idx, b), -1,
-						&bp_offset, &bp,
+						&bp_pos, &bp,
 						BTREE_ITER_NOPRESERVE);
 		if (ret) {
 			ob = ERR_PTR(ret);
 			goto err;
 		}
 
-		if (bp_offset != U64_MAX) {
+		if (!bkey_eq(bp_pos, POS_MAX)) {
 			/*
 			 * Bucket may have data in it - we don't call
 			 * bc2h_trans_inconnsistent() because fsck hasn't
@@ -358,42 +375,17 @@ static struct open_bucket *try_alloc_bucket(struct btree_trans *trans, struct bc
 		}
 	}
 
-	ob = __try_alloc_bucket(c, ca, b, reserve, &a, s, cl);
+	ob = __try_alloc_bucket(c, ca, b, watermark, a, s, cl);
 	if (!ob)
 		iter.path->preserve = false;
 err:
-	set_btree_iter_dontneed(&iter);
+	if (iter.trans && iter.path)
+		set_btree_iter_dontneed(&iter);
 	bch2_trans_iter_exit(trans, &iter);
 	printbuf_exit(&buf);
 	return ob;
 }
 
-static struct open_bucket *try_alloc_partial_bucket(struct bch_fs *c, struct bch_dev *ca,
-						    enum alloc_reserve reserve)
-{
-	struct open_bucket *ob;
-	int i;
-
-	spin_lock(&c->freelist_lock);
-
-	for (i = ca->open_buckets_partial_nr - 1; i >= 0; --i) {
-		ob = c->open_buckets + ca->open_buckets_partial[i];
-
-		if (reserve <= ob->alloc_reserve) {
-			array_remove_item(ca->open_buckets_partial,
-					  ca->open_buckets_partial_nr,
-					  i);
-			ob->on_partial_list = false;
-			ob->alloc_reserve = reserve;
-			spin_unlock(&c->freelist_lock);
-			return ob;
-		}
-	}
-
-	spin_unlock(&c->freelist_lock);
-	return NULL;
-}
-
 /*
  * This path is for before the freespace btree is initialized:
  *
@@ -403,21 +395,32 @@ static struct open_bucket *try_alloc_partial_bucket(struct bch_fs *c, struct bch
 static noinline struct open_bucket *
 bch2_bucket_alloc_early(struct btree_trans *trans,
 			struct bch_dev *ca,
-			enum alloc_reserve reserve,
+			enum bch_watermark watermark,
 			struct bucket_alloc_state *s,
 			struct closure *cl)
 {
-	struct btree_iter iter;
-	struct bkey_s_c k;
+	struct btree_iter iter, citer;
+	struct bkey_s_c k, ck;
 	struct open_bucket *ob = NULL;
+	u64 first_bucket = max_t(u64, ca->mi.first_bucket, ca->new_fs_bucket_idx);
+	u64 alloc_start = max(first_bucket, READ_ONCE(ca->alloc_cursor));
+	u64 alloc_cursor = alloc_start;
 	int ret;
 
-	s->cur_bucket = max_t(u64, s->cur_bucket, ca->mi.first_bucket);
-	s->cur_bucket = max_t(u64, s->cur_bucket, ca->new_fs_bucket_idx);
-
-	for_each_btree_key_norestart(trans, iter, BTREE_ID_alloc, POS(ca->dev_idx, s->cur_bucket),
+	/*
+	 * Scan with an uncached iterator to avoid polluting the key cache. An
+	 * uncached iter will return a cached key if one exists, but if not
+	 * there is no other underlying protection for the associated key cache
+	 * slot. To avoid racing bucket allocations, look up the cached key slot
+	 * of any likely allocation candidate before attempting to proceed with
+	 * the allocation. This provides proper exclusion on the associated
+	 * bucket.
+	 */
+again:
+	for_each_btree_key_norestart(trans, iter, BTREE_ID_alloc, POS(ca->dev_idx, alloc_cursor),
 			   BTREE_ITER_SLOTS, k, ret) {
-		struct bch_alloc_v4 a;
+		struct bch_alloc_v4 a_convert;
+		const struct bch_alloc_v4 *a;
 
 		if (bkey_ge(k.k->p, POS(ca->dev_idx, ca->mi.nbuckets)))
 			break;
@@ -426,60 +429,82 @@ bch2_bucket_alloc_early(struct btree_trans *trans,
 		    is_superblock_bucket(ca, k.k->p.offset))
 			continue;
 
-		bch2_alloc_to_v4(k, &a);
-
-		if (a.data_type != BCH_DATA_free)
+		a = bch2_alloc_to_v4(k, &a_convert);
+		if (a->data_type != BCH_DATA_free)
 			continue;
 
+		/* now check the cached key to serialize concurrent allocs of the bucket */
+		ck = bch2_bkey_get_iter(trans, &citer, BTREE_ID_alloc, k.k->p, BTREE_ITER_CACHED);
+		ret = bkey_err(ck);
+		if (ret)
+			break;
+
+		a = bch2_alloc_to_v4(ck, &a_convert);
+		if (a->data_type != BCH_DATA_free)
+			goto next;
+
 		s->buckets_seen++;
 
-		ob = __try_alloc_bucket(trans->c, ca, k.k->p.offset, reserve, &a, s, cl);
+		ob = __try_alloc_bucket(trans->c, ca, k.k->p.offset, watermark, a, s, cl);
+next:
+		citer.path->preserve = false;
+		bch2_trans_iter_exit(trans, &citer);
 		if (ob)
 			break;
 	}
 	bch2_trans_iter_exit(trans, &iter);
 
-	s->cur_bucket = iter.pos.offset;
+	alloc_cursor = iter.pos.offset;
+	ca->alloc_cursor = alloc_cursor;
+
+	if (!ob && ret)
+		ob = ERR_PTR(ret);
 
-	return ob ?: ERR_PTR(ret ?: -BCH_ERR_no_buckets_found);
+	if (!ob && alloc_start > first_bucket) {
+		alloc_cursor = alloc_start = first_bucket;
+		goto again;
+	}
+
+	return ob;
 }
 
 static struct open_bucket *bch2_bucket_alloc_freelist(struct btree_trans *trans,
 						   struct bch_dev *ca,
-						   enum alloc_reserve reserve,
+						   enum bch_watermark watermark,
 						   struct bucket_alloc_state *s,
 						   struct closure *cl)
 {
 	struct btree_iter iter;
 	struct bkey_s_c k;
 	struct open_bucket *ob = NULL;
+	u64 alloc_start = max_t(u64, ca->mi.first_bucket, READ_ONCE(ca->alloc_cursor));
+	u64 alloc_cursor = alloc_start;
 	int ret;
 
 	BUG_ON(ca->new_fs_bucket_idx);
-
-	/*
-	 * XXX:
-	 * On transaction restart, we'd like to restart from the bucket we were
-	 * at previously
-	 */
+again:
 	for_each_btree_key_norestart(trans, iter, BTREE_ID_freespace,
-				     POS(ca->dev_idx, s->cur_bucket), 0, k, ret) {
+				     POS(ca->dev_idx, alloc_cursor), 0, k, ret) {
 		if (k.k->p.inode != ca->dev_idx)
 			break;
 
-		for (s->cur_bucket = max(s->cur_bucket, bkey_start_offset(k.k));
-		     s->cur_bucket < k.k->p.offset;
-		     s->cur_bucket++) {
+		for (alloc_cursor = max(alloc_cursor, bkey_start_offset(k.k));
+		     alloc_cursor < k.k->p.offset;
+		     alloc_cursor++) {
 			ret = btree_trans_too_many_iters(trans);
-			if (ret)
+			if (ret) {
+				ob = ERR_PTR(ret);
 				break;
+			}
 
 			s->buckets_seen++;
 
-			ob = try_alloc_bucket(trans, ca, reserve,
-					      s->cur_bucket, s, k, cl);
-			if (ob)
+			ob = try_alloc_bucket(trans, ca, watermark,
+					      alloc_cursor, s, k, cl);
+			if (ob) {
+				iter.path->preserve = false;
 				break;
+			}
 		}
 
 		if (ob || ret)
@@ -487,31 +512,44 @@ static struct open_bucket *bch2_bucket_alloc_freelist(struct btree_trans *trans,
 	}
 	bch2_trans_iter_exit(trans, &iter);
 
-	return ob ?: ERR_PTR(ret);
+	ca->alloc_cursor = alloc_cursor;
+
+	if (!ob && ret)
+		ob = ERR_PTR(ret);
+
+	if (!ob && alloc_start > ca->mi.first_bucket) {
+		alloc_cursor = alloc_start = ca->mi.first_bucket;
+		goto again;
+	}
+
+	return ob;
 }
 
 /**
- * bch_bucket_alloc - allocate a single bucket from a specific device
+ * bch2_bucket_alloc_trans - allocate a single bucket from a specific device
+ * @trans:	transaction object
+ * @ca:		device to allocate from
+ * @watermark:	how important is this allocation?
+ * @cl:		if not NULL, closure to be used to wait if buckets not available
+ * @usage:	for secondarily also returning the current device usage
  *
- * Returns index of bucket on success, 0 on failure
+ * Returns:	an open_bucket on success, or an ERR_PTR() on failure.
  */
 static struct open_bucket *bch2_bucket_alloc_trans(struct btree_trans *trans,
 				      struct bch_dev *ca,
-				      enum alloc_reserve reserve,
-				      bool may_alloc_partial,
+				      enum bch_watermark watermark,
 				      struct closure *cl,
 				      struct bch_dev_usage *usage)
 {
 	struct bch_fs *c = trans->c;
 	struct open_bucket *ob = NULL;
-	bool freespace_initialized = READ_ONCE(ca->mi.freespace_initialized);
-	u64 start = freespace_initialized ? 0 : ca->bucket_alloc_trans_early_cursor;
+	bool freespace = READ_ONCE(ca->mi.freespace_initialized);
 	u64 avail;
-	struct bucket_alloc_state s = { .cur_bucket = start };
+	struct bucket_alloc_state s = { 0 };
 	bool waiting = false;
 again:
 	bch2_dev_usage_read_fast(ca, usage);
-	avail = dev_buckets_free(ca, *usage, reserve);
+	avail = dev_buckets_free(ca, *usage, watermark);
 
 	if (usage->d[BCH_DATA_need_discard].buckets > avail)
 		bch2_do_discards(c);
@@ -538,37 +576,37 @@ again:
 
 	if (waiting)
 		closure_wake_up(&c->freelist_wait);
-
-	if (may_alloc_partial) {
-		ob = try_alloc_partial_bucket(c, ca, reserve);
-		if (ob)
-			return ob;
-	}
-
-	ob = likely(ca->mi.freespace_initialized)
-		? bch2_bucket_alloc_freelist(trans, ca, reserve, &s, cl)
-		: bch2_bucket_alloc_early(trans, ca, reserve, &s, cl);
+alloc:
+	ob = likely(freespace)
+		? bch2_bucket_alloc_freelist(trans, ca, watermark, &s, cl)
+		: bch2_bucket_alloc_early(trans, ca, watermark, &s, cl);
 
 	if (s.skipped_need_journal_commit * 2 > avail)
 		bch2_journal_flush_async(&c->journal, NULL);
 
-	if (!ob && !freespace_initialized && start) {
-		start = s.cur_bucket = 0;
-		goto again;
+	if (!ob && freespace && c->curr_recovery_pass <= BCH_RECOVERY_PASS_check_alloc_info) {
+		freespace = false;
+		goto alloc;
 	}
-
-	if (!freespace_initialized)
-		ca->bucket_alloc_trans_early_cursor = s.cur_bucket;
 err:
 	if (!ob)
 		ob = ERR_PTR(-BCH_ERR_no_buckets_found);
 
 	if (!IS_ERR(ob))
-		trace_and_count(c, bucket_alloc, ca, bch2_alloc_reserves[reserve],
-				may_alloc_partial, ob->bucket);
+		trace_and_count(c, bucket_alloc, ca,
+				bch2_watermarks[watermark],
+				ob->bucket,
+				usage->d[BCH_DATA_free].buckets,
+				avail,
+				bch2_copygc_wait_amount(c),
+				c->copygc_wait - atomic64_read(&c->io_clock[WRITE].now),
+				&s,
+				cl == NULL,
+				"");
 	else if (!bch2_err_matches(PTR_ERR(ob), BCH_ERR_transaction_restart))
-		trace_and_count(c, bucket_alloc_fail,
-				ca, bch2_alloc_reserves[reserve],
+		trace_and_count(c, bucket_alloc_fail, ca,
+				bch2_watermarks[watermark],
+				0,
 				usage->d[BCH_DATA_free].buckets,
 				avail,
 				bch2_copygc_wait_amount(c),
@@ -581,16 +619,15 @@ err:
 }
 
 struct open_bucket *bch2_bucket_alloc(struct bch_fs *c, struct bch_dev *ca,
-				      enum alloc_reserve reserve,
-				      bool may_alloc_partial,
+				      enum bch_watermark watermark,
 				      struct closure *cl)
 {
 	struct bch_dev_usage usage;
 	struct open_bucket *ob;
 
 	bch2_trans_do(c, NULL, NULL, 0,
-		      PTR_ERR_OR_ZERO(ob = bch2_bucket_alloc_trans(&trans, ca, reserve,
-							may_alloc_partial, cl, &usage)));
+		      PTR_ERR_OR_ZERO(ob = bch2_bucket_alloc_trans(trans, ca, watermark,
+							cl, &usage)));
 	return ob;
 }
 
@@ -622,7 +659,7 @@ static inline void bch2_dev_stripe_increment_inlined(struct bch_dev *ca,
 			       struct bch_dev_usage *usage)
 {
 	u64 *v = stripe->next_alloc + ca->dev_idx;
-	u64 free_space = dev_buckets_available(ca, RESERVE_none);
+	u64 free_space = dev_buckets_available(ca, BCH_WATERMARK_normal);
 	u64 free_space_inv = free_space
 		? div64_u64(1ULL << 48, free_space)
 		: 1ULL << 48;
@@ -647,12 +684,10 @@ void bch2_dev_stripe_increment(struct bch_dev *ca,
 	bch2_dev_stripe_increment_inlined(ca, stripe, &usage);
 }
 
-#define BUCKET_MAY_ALLOC_PARTIAL	(1 << 0)
-#define BUCKET_ALLOC_USE_DURABILITY	(1 << 1)
-
-static void add_new_bucket(struct bch_fs *c,
+static int add_new_bucket(struct bch_fs *c,
 			   struct open_buckets *ptrs,
 			   struct bch_devs_mask *devs_may_alloc,
+			   unsigned nr_replicas,
 			   unsigned *nr_effective,
 			   bool *have_cache,
 			   unsigned flags,
@@ -661,23 +696,31 @@ static void add_new_bucket(struct bch_fs *c,
 	unsigned durability =
 		bch_dev_bkey_exists(c, ob->dev)->mi.durability;
 
+	BUG_ON(*nr_effective >= nr_replicas);
+
 	__clear_bit(ob->dev, devs_may_alloc->d);
-	*nr_effective	+= (flags & BUCKET_ALLOC_USE_DURABILITY)
-		? durability : 1;
+	*nr_effective	+= durability;
 	*have_cache	|= !durability;
 
 	ob_push(c, ptrs, ob);
+
+	if (*nr_effective >= nr_replicas)
+		return 1;
+	if (ob->ec)
+		return 1;
+	return 0;
 }
 
-static int bch2_bucket_alloc_set_trans(struct btree_trans *trans,
+int bch2_bucket_alloc_set_trans(struct btree_trans *trans,
 		      struct open_buckets *ptrs,
 		      struct dev_stripe_state *stripe,
 		      struct bch_devs_mask *devs_may_alloc,
 		      unsigned nr_replicas,
 		      unsigned *nr_effective,
 		      bool *have_cache,
-		      enum alloc_reserve reserve,
 		      unsigned flags,
+		      enum bch_data_type data_type,
+		      enum bch_watermark watermark,
 		      struct closure *cl)
 {
 	struct bch_fs *c = trans->c;
@@ -710,8 +753,7 @@ static int bch2_bucket_alloc_set_trans(struct btree_trans *trans,
 			continue;
 		}
 
-		ob = bch2_bucket_alloc_trans(trans, ca, reserve,
-				flags & BUCKET_MAY_ALLOC_PARTIAL, cl, &usage);
+		ob = bch2_bucket_alloc_trans(trans, ca, watermark, cl, &usage);
 		if (!IS_ERR(ob))
 			bch2_dev_stripe_increment_inlined(ca, stripe, &usage);
 		percpu_ref_put(&ca->ref);
@@ -723,10 +765,11 @@ static int bch2_bucket_alloc_set_trans(struct btree_trans *trans,
 			continue;
 		}
 
-		add_new_bucket(c, ptrs, devs_may_alloc,
-			       nr_effective, have_cache, flags, ob);
+		ob->data_type = data_type;
 
-		if (*nr_effective >= nr_replicas) {
+		if (add_new_bucket(c, ptrs, devs_may_alloc,
+				   nr_replicas, nr_effective,
+				   have_cache, flags, ob)) {
 			ret = 0;
 			break;
 		}
@@ -735,24 +778,6 @@ static int bch2_bucket_alloc_set_trans(struct btree_trans *trans,
 	return ret;
 }
 
-int bch2_bucket_alloc_set(struct bch_fs *c,
-		      struct open_buckets *ptrs,
-		      struct dev_stripe_state *stripe,
-		      struct bch_devs_mask *devs_may_alloc,
-		      unsigned nr_replicas,
-		      unsigned *nr_effective,
-		      bool *have_cache,
-		      enum alloc_reserve reserve,
-		      unsigned flags,
-		      struct closure *cl)
-{
-	return bch2_trans_do(c, NULL, NULL, 0,
-		      bch2_bucket_alloc_set_trans(&trans, ptrs, stripe,
-					      devs_may_alloc, nr_replicas,
-					      nr_effective, have_cache, reserve,
-					      flags, cl));
-}
-
 /* Allocate from stripes: */
 
 /*
@@ -761,26 +786,24 @@ int bch2_bucket_alloc_set(struct bch_fs *c,
  * it's to a device we don't want:
  */
 
-static int bucket_alloc_from_stripe(struct bch_fs *c,
+static int bucket_alloc_from_stripe(struct btree_trans *trans,
 			 struct open_buckets *ptrs,
 			 struct write_point *wp,
 			 struct bch_devs_mask *devs_may_alloc,
 			 u16 target,
-			 unsigned erasure_code,
 			 unsigned nr_replicas,
 			 unsigned *nr_effective,
 			 bool *have_cache,
+			 enum bch_watermark watermark,
 			 unsigned flags,
 			 struct closure *cl)
 {
+	struct bch_fs *c = trans->c;
 	struct dev_alloc_list devs_sorted;
 	struct ec_stripe_head *h;
 	struct open_bucket *ob;
-	struct bch_dev *ca;
 	unsigned i, ec_idx;
-
-	if (!erasure_code)
-		return 0;
+	int ret = 0;
 
 	if (nr_replicas < 2)
 		return 0;
@@ -788,11 +811,9 @@ static int bucket_alloc_from_stripe(struct bch_fs *c,
 	if (ec_open_bucket(c, ptrs))
 		return 0;
 
-	h = bch2_ec_stripe_head_get(c, target, 0, nr_replicas - 1,
-				    wp == &c->copygc_write_point,
-				    cl);
+	h = bch2_ec_stripe_head_get(trans, target, 0, nr_replicas - 1, watermark, cl);
 	if (IS_ERR(h))
-		return -PTR_ERR(h);
+		return PTR_ERR(h);
 	if (!h)
 		return 0;
 
@@ -810,63 +831,132 @@ static int bucket_alloc_from_stripe(struct bch_fs *c,
 		}
 	goto out_put_head;
 got_bucket:
-	ca = bch_dev_bkey_exists(c, ob->dev);
-
 	ob->ec_idx	= ec_idx;
 	ob->ec		= h->s;
+	ec_stripe_new_get(h->s, STRIPE_REF_io);
 
-	add_new_bucket(c, ptrs, devs_may_alloc,
-		       nr_effective, have_cache, flags, ob);
-	atomic_inc(&h->s->pin);
+	ret = add_new_bucket(c, ptrs, devs_may_alloc,
+			     nr_replicas, nr_effective,
+			     have_cache, flags, ob);
 out_put_head:
 	bch2_ec_stripe_head_put(c, h);
-	return 0;
+	return ret;
 }
 
 /* Sector allocator */
 
-static void get_buckets_from_writepoint(struct bch_fs *c,
-					struct open_buckets *ptrs,
-					struct write_point *wp,
-					struct bch_devs_mask *devs_may_alloc,
-					unsigned nr_replicas,
-					unsigned *nr_effective,
-					bool *have_cache,
-					unsigned flags,
-					bool need_ec)
+static bool want_bucket(struct bch_fs *c,
+			struct write_point *wp,
+			struct bch_devs_mask *devs_may_alloc,
+			bool *have_cache, bool ec,
+			struct open_bucket *ob)
+{
+	struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
+
+	if (!test_bit(ob->dev, devs_may_alloc->d))
+		return false;
+
+	if (ob->data_type != wp->data_type)
+		return false;
+
+	if (!ca->mi.durability &&
+	    (wp->data_type == BCH_DATA_btree || ec || *have_cache))
+		return false;
+
+	if (ec != (ob->ec != NULL))
+		return false;
+
+	return true;
+}
+
+static int bucket_alloc_set_writepoint(struct bch_fs *c,
+				       struct open_buckets *ptrs,
+				       struct write_point *wp,
+				       struct bch_devs_mask *devs_may_alloc,
+				       unsigned nr_replicas,
+				       unsigned *nr_effective,
+				       bool *have_cache,
+				       bool ec, unsigned flags)
 {
 	struct open_buckets ptrs_skip = { .nr = 0 };
 	struct open_bucket *ob;
 	unsigned i;
+	int ret = 0;
 
 	open_bucket_for_each(c, &wp->ptrs, ob, i) {
-		struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
-
-		if (*nr_effective < nr_replicas &&
-		    test_bit(ob->dev, devs_may_alloc->d) &&
-		    (ca->mi.durability ||
-		     (wp->data_type == BCH_DATA_user && !*have_cache)) &&
-		    (ob->ec || !need_ec)) {
-			add_new_bucket(c, ptrs, devs_may_alloc,
-				       nr_effective, have_cache,
-				       flags, ob);
-		} else {
+		if (!ret && want_bucket(c, wp, devs_may_alloc,
+					have_cache, ec, ob))
+			ret = add_new_bucket(c, ptrs, devs_may_alloc,
+				       nr_replicas, nr_effective,
+				       have_cache, flags, ob);
+		else
 			ob_push(c, &ptrs_skip, ob);
-		}
 	}
 	wp->ptrs = ptrs_skip;
+
+	return ret;
 }
 
-static int open_bucket_add_buckets(struct btree_trans *trans,
+static int bucket_alloc_set_partial(struct bch_fs *c,
+				    struct open_buckets *ptrs,
+				    struct write_point *wp,
+				    struct bch_devs_mask *devs_may_alloc,
+				    unsigned nr_replicas,
+				    unsigned *nr_effective,
+				    bool *have_cache, bool ec,
+				    enum bch_watermark watermark,
+				    unsigned flags)
+{
+	int i, ret = 0;
+
+	if (!c->open_buckets_partial_nr)
+		return 0;
+
+	spin_lock(&c->freelist_lock);
+
+	if (!c->open_buckets_partial_nr)
+		goto unlock;
+
+	for (i = c->open_buckets_partial_nr - 1; i >= 0; --i) {
+		struct open_bucket *ob = c->open_buckets + c->open_buckets_partial[i];
+
+		if (want_bucket(c, wp, devs_may_alloc, have_cache, ec, ob)) {
+			struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
+			struct bch_dev_usage usage;
+			u64 avail;
+
+			bch2_dev_usage_read_fast(ca, &usage);
+			avail = dev_buckets_free(ca, usage, watermark);
+			if (!avail)
+				continue;
+
+			array_remove_item(c->open_buckets_partial,
+					  c->open_buckets_partial_nr,
+					  i);
+			ob->on_partial_list = false;
+
+			ret = add_new_bucket(c, ptrs, devs_may_alloc,
+					     nr_replicas, nr_effective,
+					     have_cache, flags, ob);
+			if (ret)
+				break;
+		}
+	}
+unlock:
+	spin_unlock(&c->freelist_lock);
+	return ret;
+}
+
+static int __open_bucket_add_buckets(struct btree_trans *trans,
 			struct open_buckets *ptrs,
 			struct write_point *wp,
 			struct bch_devs_list *devs_have,
 			u16 target,
-			unsigned erasure_code,
+			bool erasure_code,
 			unsigned nr_replicas,
 			unsigned *nr_effective,
 			bool *have_cache,
-			enum alloc_reserve reserve,
+			enum bch_watermark watermark,
 			unsigned flags,
 			struct closure *_cl)
 {
@@ -874,12 +964,10 @@ static int open_bucket_add_buckets(struct btree_trans *trans,
 	struct bch_devs_mask devs;
 	struct open_bucket *ob;
 	struct closure *cl = NULL;
-	int ret;
 	unsigned i;
+	int ret;
 
-	rcu_read_lock();
 	devs = target_rw_devs(c, wp->data_type, target);
-	rcu_read_unlock();
 
 	/* Don't allocate from devices we already have pointers to: */
 	for (i = 0; i < devs_have->nr; i++)
@@ -888,93 +976,191 @@ static int open_bucket_add_buckets(struct btree_trans *trans,
 	open_bucket_for_each(c, ptrs, ob, i)
 		__clear_bit(ob->dev, devs.d);
 
-	if (erasure_code) {
-		if (!ec_open_bucket(c, ptrs)) {
-			get_buckets_from_writepoint(c, ptrs, wp, &devs,
-						    nr_replicas, nr_effective,
-						    have_cache, flags, true);
-			if (*nr_effective >= nr_replicas)
-				return 0;
-		}
+	if (erasure_code && ec_open_bucket(c, ptrs))
+		return 0;
 
-		if (!ec_open_bucket(c, ptrs)) {
-			ret = bucket_alloc_from_stripe(c, ptrs, wp, &devs,
-						 target, erasure_code,
-						 nr_replicas, nr_effective,
-						 have_cache, flags, _cl);
-			if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
-			    bch2_err_matches(ret, BCH_ERR_freelist_empty) ||
-			    bch2_err_matches(ret, BCH_ERR_open_buckets_empty))
-				return ret;
-			if (*nr_effective >= nr_replicas)
-				return 0;
-		}
-	}
+	ret = bucket_alloc_set_writepoint(c, ptrs, wp, &devs,
+				 nr_replicas, nr_effective,
+				 have_cache, erasure_code, flags);
+	if (ret)
+		return ret;
 
-	get_buckets_from_writepoint(c, ptrs, wp, &devs,
-				    nr_replicas, nr_effective,
-				    have_cache, flags, false);
-	if (*nr_effective >= nr_replicas)
-		return 0;
+	ret = bucket_alloc_set_partial(c, ptrs, wp, &devs,
+				 nr_replicas, nr_effective,
+				 have_cache, erasure_code, watermark, flags);
+	if (ret)
+		return ret;
 
+	if (erasure_code) {
+		ret = bucket_alloc_from_stripe(trans, ptrs, wp, &devs,
+					 target,
+					 nr_replicas, nr_effective,
+					 have_cache,
+					 watermark, flags, _cl);
+	} else {
 retry_blocking:
-	/*
-	 * Try nonblocking first, so that if one device is full we'll try from
-	 * other devices:
-	 */
-	ret = bch2_bucket_alloc_set_trans(trans, ptrs, &wp->stripe, &devs,
-				nr_replicas, nr_effective, have_cache,
-				reserve, flags, cl);
-	if (ret &&
-	    !bch2_err_matches(ret, BCH_ERR_transaction_restart) &&
-	    !bch2_err_matches(ret, BCH_ERR_insufficient_devices) &&
-	    !cl && _cl) {
-		cl = _cl;
-		goto retry_blocking;
+		/*
+		 * Try nonblocking first, so that if one device is full we'll try from
+		 * other devices:
+		 */
+		ret = bch2_bucket_alloc_set_trans(trans, ptrs, &wp->stripe, &devs,
+					nr_replicas, nr_effective, have_cache,
+					flags, wp->data_type, watermark, cl);
+		if (ret &&
+		    !bch2_err_matches(ret, BCH_ERR_transaction_restart) &&
+		    !bch2_err_matches(ret, BCH_ERR_insufficient_devices) &&
+		    !cl && _cl) {
+			cl = _cl;
+			goto retry_blocking;
+		}
 	}
 
 	return ret;
 }
 
-void bch2_open_buckets_stop_dev(struct bch_fs *c, struct bch_dev *ca,
-				struct open_buckets *obs)
+static int open_bucket_add_buckets(struct btree_trans *trans,
+			struct open_buckets *ptrs,
+			struct write_point *wp,
+			struct bch_devs_list *devs_have,
+			u16 target,
+			unsigned erasure_code,
+			unsigned nr_replicas,
+			unsigned *nr_effective,
+			bool *have_cache,
+			enum bch_watermark watermark,
+			unsigned flags,
+			struct closure *cl)
 {
-	struct open_buckets ptrs = { .nr = 0 };
-	struct open_bucket *ob, *ob2;
-	unsigned i, j;
+	int ret;
 
-	open_bucket_for_each(c, obs, ob, i) {
-		bool drop = !ca || ob->dev == ca->dev_idx;
+	if (erasure_code) {
+		ret = __open_bucket_add_buckets(trans, ptrs, wp,
+				devs_have, target, erasure_code,
+				nr_replicas, nr_effective, have_cache,
+				watermark, flags, cl);
+		if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
+		    bch2_err_matches(ret, BCH_ERR_operation_blocked) ||
+		    bch2_err_matches(ret, BCH_ERR_freelist_empty) ||
+		    bch2_err_matches(ret, BCH_ERR_open_buckets_empty))
+			return ret;
+		if (*nr_effective >= nr_replicas)
+			return 0;
+	}
+
+	ret = __open_bucket_add_buckets(trans, ptrs, wp,
+			devs_have, target, false,
+			nr_replicas, nr_effective, have_cache,
+			watermark, flags, cl);
+	return ret < 0 ? ret : 0;
+}
+
+/**
+ * should_drop_bucket - check if this is open_bucket should go away
+ * @ob:		open_bucket to predicate on
+ * @c:		filesystem handle
+ * @ca:		if set, we're killing buckets for a particular device
+ * @ec:		if true, we're shutting down erasure coding and killing all ec
+ *		open_buckets
+ *		otherwise, return true
+ * Returns: true if we should kill this open_bucket
+ *
+ * We're killing open_buckets because we're shutting down a device, erasure
+ * coding, or the entire filesystem - check if this open_bucket matches:
+ */
+static bool should_drop_bucket(struct open_bucket *ob, struct bch_fs *c,
+			       struct bch_dev *ca, bool ec)
+{
+	if (ec) {
+		return ob->ec != NULL;
+	} else if (ca) {
+		bool drop = ob->dev == ca->dev_idx;
+		struct open_bucket *ob2;
+		unsigned i;
 
 		if (!drop && ob->ec) {
+			unsigned nr_blocks;
+
 			mutex_lock(&ob->ec->lock);
-			for (j = 0; j < ob->ec->new_stripe.key.v.nr_blocks; j++) {
-				if (!ob->ec->blocks[j])
+			nr_blocks = bkey_i_to_stripe(&ob->ec->new_stripe.key)->v.nr_blocks;
+
+			for (i = 0; i < nr_blocks; i++) {
+				if (!ob->ec->blocks[i])
 					continue;
 
-				ob2 = c->open_buckets + ob->ec->blocks[j];
+				ob2 = c->open_buckets + ob->ec->blocks[i];
 				drop |= ob2->dev == ca->dev_idx;
 			}
 			mutex_unlock(&ob->ec->lock);
 		}
 
-		if (drop)
-			bch2_open_bucket_put(c, ob);
-		else
-			ob_push(c, &ptrs, ob);
+		return drop;
+	} else {
+		return true;
 	}
-
-	*obs = ptrs;
 }
 
-void bch2_writepoint_stop(struct bch_fs *c, struct bch_dev *ca,
-			  struct write_point *wp)
+static void bch2_writepoint_stop(struct bch_fs *c, struct bch_dev *ca,
+				 bool ec, struct write_point *wp)
 {
+	struct open_buckets ptrs = { .nr = 0 };
+	struct open_bucket *ob;
+	unsigned i;
+
 	mutex_lock(&wp->lock);
-	bch2_open_buckets_stop_dev(c, ca, &wp->ptrs);
+	open_bucket_for_each(c, &wp->ptrs, ob, i)
+		if (should_drop_bucket(ob, c, ca, ec))
+			bch2_open_bucket_put(c, ob);
+		else
+			ob_push(c, &ptrs, ob);
+	wp->ptrs = ptrs;
 	mutex_unlock(&wp->lock);
 }
 
+void bch2_open_buckets_stop(struct bch_fs *c, struct bch_dev *ca,
+			    bool ec)
+{
+	unsigned i;
+
+	/* Next, close write points that point to this device... */
+	for (i = 0; i < ARRAY_SIZE(c->write_points); i++)
+		bch2_writepoint_stop(c, ca, ec, &c->write_points[i]);
+
+	bch2_writepoint_stop(c, ca, ec, &c->copygc_write_point);
+	bch2_writepoint_stop(c, ca, ec, &c->rebalance_write_point);
+	bch2_writepoint_stop(c, ca, ec, &c->btree_write_point);
+
+	mutex_lock(&c->btree_reserve_cache_lock);
+	while (c->btree_reserve_cache_nr) {
+		struct btree_alloc *a =
+			&c->btree_reserve_cache[--c->btree_reserve_cache_nr];
+
+		bch2_open_buckets_put(c, &a->ob);
+	}
+	mutex_unlock(&c->btree_reserve_cache_lock);
+
+	spin_lock(&c->freelist_lock);
+	i = 0;
+	while (i < c->open_buckets_partial_nr) {
+		struct open_bucket *ob =
+			c->open_buckets + c->open_buckets_partial[i];
+
+		if (should_drop_bucket(ob, c, ca, ec)) {
+			--c->open_buckets_partial_nr;
+			swap(c->open_buckets_partial[i],
+			     c->open_buckets_partial[c->open_buckets_partial_nr]);
+			ob->on_partial_list = false;
+			spin_unlock(&c->freelist_lock);
+			bch2_open_bucket_put(c, ob);
+			spin_lock(&c->freelist_lock);
+		} else {
+			i++;
+		}
+	}
+	spin_unlock(&c->freelist_lock);
+
+	bch2_ec_stop_dev(c, ca);
+}
+
 static inline struct hlist_head *writepoint_hash(struct bch_fs *c,
 						 unsigned long write_point)
 {
@@ -1020,10 +1206,12 @@ static bool try_increase_writepoints(struct bch_fs *c)
 	return true;
 }
 
-static bool try_decrease_writepoints(struct bch_fs *c,
-				     unsigned old_nr)
+static bool try_decrease_writepoints(struct btree_trans *trans, unsigned old_nr)
 {
+	struct bch_fs *c = trans->c;
 	struct write_point *wp;
+	struct open_bucket *ob;
+	unsigned i;
 
 	mutex_lock(&c->write_points_hash_lock);
 	if (c->write_points_nr < old_nr) {
@@ -1042,19 +1230,14 @@ static bool try_decrease_writepoints(struct bch_fs *c,
 	hlist_del_rcu(&wp->node);
 	mutex_unlock(&c->write_points_hash_lock);
 
-	bch2_writepoint_stop(c, NULL, wp);
+	bch2_trans_mutex_lock_norelock(trans, &wp->lock);
+	open_bucket_for_each(c, &wp->ptrs, ob, i)
+		open_bucket_free_unused(c, ob);
+	wp->ptrs.nr = 0;
+	mutex_unlock(&wp->lock);
 	return true;
 }
 
-static void bch2_trans_mutex_lock(struct btree_trans *trans,
-				  struct mutex *lock)
-{
-	if (!mutex_trylock(lock)) {
-		bch2_trans_unlock(trans);
-		mutex_lock(lock);
-	}
-}
-
 static struct write_point *writepoint_find(struct btree_trans *trans,
 					   unsigned long write_point)
 {
@@ -1064,7 +1247,7 @@ static struct write_point *writepoint_find(struct btree_trans *trans,
 
 	if (!(write_point & 1UL)) {
 		wp = (struct write_point *) write_point;
-		bch2_trans_mutex_lock(trans, &wp->lock);
+		bch2_trans_mutex_lock_norelock(trans, &wp->lock);
 		return wp;
 	}
 
@@ -1073,7 +1256,7 @@ restart_find:
 	wp = __writepoint_find(head, write_point);
 	if (wp) {
 lock_wp:
-		bch2_trans_mutex_lock(trans, &wp->lock);
+		bch2_trans_mutex_lock_norelock(trans, &wp->lock);
 		if (wp->write_point == write_point)
 			goto out;
 		mutex_unlock(&wp->lock);
@@ -1086,8 +1269,8 @@ restart_find_oldest:
 		if (!oldest || time_before64(wp->last_used, oldest->last_used))
 			oldest = wp;
 
-	bch2_trans_mutex_lock(trans, &oldest->lock);
-	bch2_trans_mutex_lock(trans, &c->write_points_hash_lock);
+	bch2_trans_mutex_lock_norelock(trans, &oldest->lock);
+	bch2_trans_mutex_lock_norelock(trans, &c->write_points_hash_lock);
 	if (oldest >= c->write_points + c->write_points_nr ||
 	    try_increase_writepoints(c)) {
 		mutex_unlock(&c->write_points_hash_lock);
@@ -1112,33 +1295,58 @@ out:
 	return wp;
 }
 
+static noinline void
+deallocate_extra_replicas(struct bch_fs *c,
+			  struct open_buckets *ptrs,
+			  struct open_buckets *ptrs_no_use,
+			  unsigned extra_replicas)
+{
+	struct open_buckets ptrs2 = { 0 };
+	struct open_bucket *ob;
+	unsigned i;
+
+	open_bucket_for_each(c, ptrs, ob, i) {
+		unsigned d = bch_dev_bkey_exists(c, ob->dev)->mi.durability;
+
+		if (d && d <= extra_replicas) {
+			extra_replicas -= d;
+			ob_push(c, ptrs_no_use, ob);
+		} else {
+			ob_push(c, &ptrs2, ob);
+		}
+	}
+
+	*ptrs = ptrs2;
+}
+
 /*
  * Get us an open_bucket we can allocate from, return with it locked:
  */
 int bch2_alloc_sectors_start_trans(struct btree_trans *trans,
-				   unsigned target,
-				   unsigned erasure_code,
-				   struct write_point_specifier write_point,
-				   struct bch_devs_list *devs_have,
-				   unsigned nr_replicas,
-				   unsigned nr_replicas_required,
-				   enum alloc_reserve reserve,
-				   unsigned flags,
-				   struct closure *cl,
-				   struct write_point **wp_ret)
+			     unsigned target,
+			     unsigned erasure_code,
+			     struct write_point_specifier write_point,
+			     struct bch_devs_list *devs_have,
+			     unsigned nr_replicas,
+			     unsigned nr_replicas_required,
+			     enum bch_watermark watermark,
+			     unsigned flags,
+			     struct closure *cl,
+			     struct write_point **wp_ret)
 {
 	struct bch_fs *c = trans->c;
 	struct write_point *wp;
 	struct open_bucket *ob;
 	struct open_buckets ptrs;
 	unsigned nr_effective, write_points_nr;
-	unsigned ob_flags = 0;
 	bool have_cache;
 	int ret;
 	int i;
 
-	if (!(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS))
-		ob_flags |= BUCKET_ALLOC_USE_DURABILITY;
+	if (!IS_ENABLED(CONFIG_BCACHEFS_ERASURE_CODING))
+		erasure_code = false;
+
+	BUG_ON(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS);
 
 	BUG_ON(!nr_replicas || !nr_replicas_required);
 retry:
@@ -1149,34 +1357,42 @@ retry:
 
 	*wp_ret = wp = writepoint_find(trans, write_point.v);
 
-	if (wp->data_type == BCH_DATA_user)
-		ob_flags |= BUCKET_MAY_ALLOC_PARTIAL;
-
 	/* metadata may not allocate on cache devices: */
 	if (wp->data_type != BCH_DATA_user)
 		have_cache = true;
 
-	if (!target || (flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)) {
-		ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
-					      target, erasure_code,
-					      nr_replicas, &nr_effective,
-					      &have_cache, reserve,
-					      ob_flags, cl);
-	} else {
+	if (target && !(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)) {
 		ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
 					      target, erasure_code,
 					      nr_replicas, &nr_effective,
-					      &have_cache, reserve,
-					      ob_flags, NULL);
+					      &have_cache, watermark,
+					      flags, NULL);
 		if (!ret ||
 		    bch2_err_matches(ret, BCH_ERR_transaction_restart))
 			goto alloc_done;
 
+		/* Don't retry from all devices if we're out of open buckets: */
+		if (bch2_err_matches(ret, BCH_ERR_open_buckets_empty))
+			goto allocate_blocking;
+
+		/*
+		 * Only try to allocate cache (durability = 0 devices) from the
+		 * specified target:
+		 */
+		have_cache = true;
+
 		ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
 					      0, erasure_code,
 					      nr_replicas, &nr_effective,
-					      &have_cache, reserve,
-					      ob_flags, cl);
+					      &have_cache, watermark,
+					      flags, cl);
+	} else {
+allocate_blocking:
+		ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
+					      target, erasure_code,
+					      nr_replicas, &nr_effective,
+					      &have_cache, watermark,
+					      flags, cl);
 	}
 alloc_done:
 	BUG_ON(!ret && nr_effective < nr_replicas);
@@ -1191,9 +1407,12 @@ alloc_done:
 	if (ret)
 		goto err;
 
+	if (nr_effective > nr_replicas)
+		deallocate_extra_replicas(c, &ptrs, &wp->ptrs, nr_effective - nr_replicas);
+
 	/* Free buckets we didn't use: */
 	open_bucket_for_each(c, &wp->ptrs, ob, i)
-		open_bucket_free_unused(c, wp, ob);
+		open_bucket_free_unused(c, ob);
 
 	wp->ptrs = ptrs;
 
@@ -1210,24 +1429,21 @@ err:
 		if (ptrs.nr < ARRAY_SIZE(ptrs.v))
 			ob_push(c, &ptrs, ob);
 		else
-			open_bucket_free_unused(c, wp, ob);
+			open_bucket_free_unused(c, ob);
 	wp->ptrs = ptrs;
 
 	mutex_unlock(&wp->lock);
 
 	if (bch2_err_matches(ret, BCH_ERR_freelist_empty) &&
-	    try_decrease_writepoints(c, write_points_nr))
+	    try_decrease_writepoints(trans, write_points_nr))
 		goto retry;
 
 	if (bch2_err_matches(ret, BCH_ERR_open_buckets_empty) ||
 	    bch2_err_matches(ret, BCH_ERR_freelist_empty))
 		return cl
-			? -EAGAIN
+			? -BCH_ERR_bucket_alloc_blocked
 			: -BCH_ERR_ENOSPC_bucket_alloc;
 
-	if (bch2_err_matches(ret, BCH_ERR_insufficient_devices))
-		return -EROFS;
-
 	return ret;
 }
 
@@ -1307,21 +1523,106 @@ void bch2_fs_allocator_foreground_init(struct bch_fs *c)
 	}
 }
 
+static void bch2_open_bucket_to_text(struct printbuf *out, struct bch_fs *c, struct open_bucket *ob)
+{
+	struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
+	unsigned data_type = ob->data_type;
+	barrier(); /* READ_ONCE() doesn't work on bitfields */
+
+	prt_printf(out, "%zu ref %u %s %u:%llu gen %u allocated %u/%u",
+		   ob - c->open_buckets,
+		   atomic_read(&ob->pin),
+		   data_type < BCH_DATA_NR ? bch2_data_types[data_type] : "invalid data type",
+		   ob->dev, ob->bucket, ob->gen,
+		   ca->mi.bucket_size - ob->sectors_free, ca->mi.bucket_size);
+	if (ob->ec)
+		prt_printf(out, " ec idx %llu", ob->ec->idx);
+	if (ob->on_partial_list)
+		prt_str(out, " partial");
+	prt_newline(out);
+}
+
 void bch2_open_buckets_to_text(struct printbuf *out, struct bch_fs *c)
 {
 	struct open_bucket *ob;
 
+	out->atomic++;
+
 	for (ob = c->open_buckets;
 	     ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
 	     ob++) {
 		spin_lock(&ob->lock);
-		if (ob->valid && !ob->on_partial_list) {
-			prt_printf(out, "%zu ref %u type %s %u:%llu:%u\n",
-			       ob - c->open_buckets,
-			       atomic_read(&ob->pin),
-			       bch2_data_types[ob->data_type],
-			       ob->dev, ob->bucket, ob->gen);
-		}
+		if (ob->valid && !ob->on_partial_list)
+			bch2_open_bucket_to_text(out, c, ob);
 		spin_unlock(&ob->lock);
 	}
+
+	--out->atomic;
+}
+
+void bch2_open_buckets_partial_to_text(struct printbuf *out, struct bch_fs *c)
+{
+	unsigned i;
+
+	out->atomic++;
+	spin_lock(&c->freelist_lock);
+
+	for (i = 0; i < c->open_buckets_partial_nr; i++)
+		bch2_open_bucket_to_text(out, c,
+				c->open_buckets + c->open_buckets_partial[i]);
+
+	spin_unlock(&c->freelist_lock);
+	--out->atomic;
+}
+
+static const char * const bch2_write_point_states[] = {
+#define x(n)	#n,
+	WRITE_POINT_STATES()
+#undef x
+	NULL
+};
+
+static void bch2_write_point_to_text(struct printbuf *out, struct bch_fs *c,
+				     struct write_point *wp)
+{
+	struct open_bucket *ob;
+	unsigned i;
+
+	prt_printf(out, "%lu: ", wp->write_point);
+	prt_human_readable_u64(out, wp->sectors_allocated);
+
+	prt_printf(out, " last wrote: ");
+	bch2_pr_time_units(out, sched_clock() - wp->last_used);
+
+	for (i = 0; i < WRITE_POINT_STATE_NR; i++) {
+		prt_printf(out, " %s: ", bch2_write_point_states[i]);
+		bch2_pr_time_units(out, wp->time[i]);
+	}
+
+	prt_newline(out);
+
+	printbuf_indent_add(out, 2);
+	open_bucket_for_each(c, &wp->ptrs, ob, i)
+		bch2_open_bucket_to_text(out, c, ob);
+	printbuf_indent_sub(out, 2);
+}
+
+void bch2_write_points_to_text(struct printbuf *out, struct bch_fs *c)
+{
+	struct write_point *wp;
+
+	prt_str(out, "Foreground write points\n");
+	for (wp = c->write_points;
+	     wp < c->write_points + ARRAY_SIZE(c->write_points);
+	     wp++)
+		bch2_write_point_to_text(out, c, wp);
+
+	prt_str(out, "Copygc write point\n");
+	bch2_write_point_to_text(out, c, &c->copygc_write_point);
+
+	prt_str(out, "Rebalance write point\n");
+	bch2_write_point_to_text(out, c, &c->rebalance_write_point);
+
+	prt_str(out, "Btree write point\n");
+	bch2_write_point_to_text(out, c, &c->btree_write_point);
 }