]> git.sesse.net Git - bcachefs-tools-debian/blobdiff - libbcachefs/alloc_foreground.c
Move c_src dirs back to toplevel
[bcachefs-tools-debian] / libbcachefs / alloc_foreground.c
index d52f30ac66bad2f536df6a99cac1f45af2de48d3..b0ff47998a9440912f940dc09e27b34e6341cb9e 100644 (file)
 #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
 };
@@ -60,11 +69,8 @@ const char * const bch2_alloc_reserves[] = {
 
 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)
+       for_each_member_device_rcu(c, ca, NULL)
                ca->alloc_cursor = 0;
        rcu_read_unlock();
 }
@@ -150,9 +156,7 @@ 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)
 {
        BUG_ON(c->open_buckets_partial_nr >=
               ARRAY_SIZE(c->open_buckets_partial));
@@ -181,14 +185,16 @@ 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;
        }
@@ -196,7 +202,7 @@ 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,
+                                             enum bch_watermark watermark,
                                              const struct bch_alloc_v4 *a,
                                              struct bucket_alloc_state *s,
                                              struct closure *cl)
@@ -226,13 +232,12 @@ 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);
 
-               if (!c->blocked_allocate_open_bucket)
-                       c->blocked_allocate_open_bucket = local_clock();
-
+               track_event_change(&c->times[BCH_TIME_blocked_allocate_open_bucket],
+                                  &c->blocked_allocate_open_bucket, true);
                spin_unlock(&c->freelist_lock);
                return ERR_PTR(-BCH_ERR_open_buckets_empty);
        }
@@ -258,26 +263,18 @@ static struct open_bucket *__try_alloc_bucket(struct bch_fs *c, struct bch_dev *
        ca->nr_open_buckets++;
        bch2_open_bucket_hash_add(c, ob);
 
-       if (c->blocked_allocate_open_bucket) {
-               bch2_time_stats_update(
-                       &c->times[BCH_TIME_blocked_allocate_open_bucket],
-                       c->blocked_allocate_open_bucket);
-               c->blocked_allocate_open_bucket = 0;
-       }
+       track_event_change(&c->times[BCH_TIME_blocked_allocate_open_bucket],
+                          &c->blocked_allocate_open_bucket, false);
 
-       if (c->blocked_allocate) {
-               bch2_time_stats_update(
-                       &c->times[BCH_TIME_blocked_allocate],
-                       c->blocked_allocate);
-               c->blocked_allocate = 0;
-       }
+       track_event_change(&c->times[BCH_TIME_blocked_allocate],
+                          &c->blocked_allocate, false);
 
        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)
@@ -303,8 +300,9 @@ 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);
@@ -314,7 +312,7 @@ static struct open_bucket *try_alloc_bucket(struct btree_trans *trans, struct bc
        a = bch2_alloc_to_v4(k, &a_convert);
 
        if (a->data_type != BCH_DATA_free) {
-               if (!test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags)) {
+               if (c->curr_recovery_pass <= BCH_RECOVERY_PASS_check_alloc_info) {
                        ob = NULL;
                        goto err;
                }
@@ -330,7 +328,7 @@ static struct open_bucket *try_alloc_bucket(struct btree_trans *trans, struct bc
        }
 
        if (genbits != (alloc_freespace_genbits(*a) >> 56) &&
-           test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags)) {
+           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);
@@ -340,22 +338,21 @@ static struct open_bucket *try_alloc_bucket(struct btree_trans *trans, struct bc
                bch2_trans_inconsistent(trans, "%s", buf.buf);
                ob = ERR_PTR(-EIO);
                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
@@ -366,11 +363,12 @@ 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;
+               set_btree_iter_dontneed(&iter);
 err:
-       set_btree_iter_dontneed(&iter);
+       if (iter.path)
+               set_btree_iter_dontneed(&iter);
        bch2_trans_iter_exit(trans, &iter);
        printbuf_exit(&buf);
        return ob;
@@ -385,16 +383,27 @@ err:
 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 alloc_start = max_t(u64, ca->mi.first_bucket, ca->new_fs_bucket_idx);
-       u64 alloc_cursor = max(alloc_start, READ_ONCE(ca->alloc_cursor));
+       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;
+
+       /*
+        * 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) {
@@ -409,25 +418,38 @@ again:
                        continue;
 
                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:
+               set_btree_iter_dontneed(&citer);
+               bch2_trans_iter_exit(trans, &citer);
                if (ob)
                        break;
        }
        bch2_trans_iter_exit(trans, &iter);
 
+       alloc_cursor = iter.pos.offset;
        ca->alloc_cursor = alloc_cursor;
 
        if (!ob && ret)
                ob = ERR_PTR(ret);
 
-       if (!ob && alloc_cursor > alloc_start) {
-               alloc_cursor = alloc_start;
+       if (!ob && alloc_start > first_bucket) {
+               alloc_cursor = alloc_start = first_bucket;
                goto again;
        }
 
@@ -436,7 +458,7 @@ again:
 
 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)
 {
@@ -465,10 +487,10 @@ again:
 
                        s->buckets_seen++;
 
-                       ob = try_alloc_bucket(trans, ca, reserve,
+                       ob = try_alloc_bucket(trans, ca, watermark,
                                              alloc_cursor, s, k, cl);
                        if (ob) {
-                               iter.path->preserve = false;
+                               set_btree_iter_dontneed(&iter);
                                break;
                        }
                }
@@ -492,13 +514,18 @@ again:
 }
 
 /**
- * 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,
+                                     enum bch_watermark watermark,
                                      struct closure *cl,
                                      struct bch_dev_usage *usage)
 {
@@ -510,7 +537,7 @@ static struct open_bucket *bch2_bucket_alloc_trans(struct btree_trans *trans,
        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);
@@ -528,8 +555,8 @@ again:
                        goto again;
                }
 
-               if (!c->blocked_allocate)
-                       c->blocked_allocate = local_clock();
+               track_event_change(&c->times[BCH_TIME_blocked_allocate],
+                                  &c->blocked_allocate, true);
 
                ob = ERR_PTR(-BCH_ERR_freelist_empty);
                goto err;
@@ -539,13 +566,13 @@ again:
                closure_wake_up(&c->freelist_wait);
 alloc:
        ob = likely(freespace)
-               ? bch2_bucket_alloc_freelist(trans, ca, reserve, &s, cl)
-               : bch2_bucket_alloc_early(trans, ca, reserve, &s, cl);
+               ? 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 && !test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags)) {
+       if (!ob && freespace && c->curr_recovery_pass <= BCH_RECOVERY_PASS_check_alloc_info) {
                freespace = false;
                goto alloc;
        }
@@ -555,7 +582,7 @@ err:
 
        if (!IS_ERR(ob))
                trace_and_count(c, bucket_alloc, ca,
-                               bch2_alloc_reserves[reserve],
+                               bch2_watermarks[watermark],
                                ob->bucket,
                                usage->d[BCH_DATA_free].buckets,
                                avail,
@@ -566,7 +593,7 @@ err:
                                "");
        else if (!bch2_err_matches(PTR_ERR(ob), BCH_ERR_transaction_restart))
                trace_and_count(c, bucket_alloc_fail, ca,
-                               bch2_alloc_reserves[reserve],
+                               bch2_watermarks[watermark],
                                0,
                                usage->d[BCH_DATA_free].buckets,
                                avail,
@@ -580,14 +607,14 @@ err:
 }
 
 struct open_bucket *bch2_bucket_alloc(struct bch_fs *c, struct bch_dev *ca,
-                                     enum alloc_reserve reserve,
+                                     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,
+                     PTR_ERR_OR_ZERO(ob = bch2_bucket_alloc_trans(trans, ca, watermark,
                                                        cl, &usage)));
        return ob;
 }
@@ -620,7 +647,7 @@ 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;
@@ -658,11 +685,9 @@ static int add_new_bucket(struct bch_fs *c,
                bch_dev_bkey_exists(c, ob->dev)->mi.durability;
 
        BUG_ON(*nr_effective >= nr_replicas);
-       BUG_ON(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS);
 
        __clear_bit(ob->dev, devs_may_alloc->d);
-       *nr_effective   += (flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)
-               ? durability : 1;
+       *nr_effective   += durability;
        *have_cache     |= !durability;
 
        ob_push(c, ptrs, ob);
@@ -683,7 +708,7 @@ int bch2_bucket_alloc_set_trans(struct btree_trans *trans,
                      bool *have_cache,
                      unsigned flags,
                      enum bch_data_type data_type,
-                     enum alloc_reserve reserve,
+                     enum bch_watermark watermark,
                      struct closure *cl)
 {
        struct bch_fs *c = trans->c;
@@ -716,7 +741,7 @@ int bch2_bucket_alloc_set_trans(struct btree_trans *trans,
                        continue;
                }
 
-               ob = bch2_bucket_alloc_trans(trans, ca, reserve, 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);
@@ -757,7 +782,7 @@ static int bucket_alloc_from_stripe(struct btree_trans *trans,
                         unsigned nr_replicas,
                         unsigned *nr_effective,
                         bool *have_cache,
-                        enum alloc_reserve reserve,
+                        enum bch_watermark watermark,
                         unsigned flags,
                         struct closure *cl)
 {
@@ -765,7 +790,6 @@ static int bucket_alloc_from_stripe(struct btree_trans *trans,
        struct dev_alloc_list devs_sorted;
        struct ec_stripe_head *h;
        struct open_bucket *ob;
-       struct bch_dev *ca;
        unsigned i, ec_idx;
        int ret = 0;
 
@@ -775,7 +799,7 @@ static int bucket_alloc_from_stripe(struct btree_trans *trans,
        if (ec_open_bucket(c, ptrs))
                return 0;
 
-       h = bch2_ec_stripe_head_get(trans, target, 0, nr_replicas - 1, reserve, cl);
+       h = bch2_ec_stripe_head_get(trans, target, 0, nr_replicas - 1, watermark, cl);
        if (IS_ERR(h))
                return PTR_ERR(h);
        if (!h)
@@ -795,8 +819,6 @@ static int bucket_alloc_from_stripe(struct btree_trans *trans,
                }
        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);
@@ -870,7 +892,7 @@ static int bucket_alloc_set_partial(struct bch_fs *c,
                                    unsigned nr_replicas,
                                    unsigned *nr_effective,
                                    bool *have_cache, bool ec,
-                                   enum alloc_reserve reserve,
+                                   enum bch_watermark watermark,
                                    unsigned flags)
 {
        int i, ret = 0;
@@ -892,7 +914,7 @@ static int bucket_alloc_set_partial(struct bch_fs *c,
                        u64 avail;
 
                        bch2_dev_usage_read_fast(ca, &usage);
-                       avail = dev_buckets_free(ca, usage, reserve);
+                       avail = dev_buckets_free(ca, usage, watermark);
                        if (!avail)
                                continue;
 
@@ -922,7 +944,7 @@ static int __open_bucket_add_buckets(struct btree_trans *trans,
                        unsigned nr_replicas,
                        unsigned *nr_effective,
                        bool *have_cache,
-                       enum alloc_reserve reserve,
+                       enum bch_watermark watermark,
                        unsigned flags,
                        struct closure *_cl)
 {
@@ -933,13 +955,11 @@ static int __open_bucket_add_buckets(struct btree_trans *trans,
        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++)
-               __clear_bit(devs_have->devs[i], devs.d);
+       darray_for_each(*devs_have, i)
+               __clear_bit(*i, devs.d);
 
        open_bucket_for_each(c, ptrs, ob, i)
                __clear_bit(ob->dev, devs.d);
@@ -955,7 +975,7 @@ static int __open_bucket_add_buckets(struct btree_trans *trans,
 
        ret = bucket_alloc_set_partial(c, ptrs, wp, &devs,
                                 nr_replicas, nr_effective,
-                                have_cache, erasure_code, reserve, flags);
+                                have_cache, erasure_code, watermark, flags);
        if (ret)
                return ret;
 
@@ -964,7 +984,7 @@ static int __open_bucket_add_buckets(struct btree_trans *trans,
                                         target,
                                         nr_replicas, nr_effective,
                                         have_cache,
-                                        reserve, flags, _cl);
+                                        watermark, flags, _cl);
        } else {
 retry_blocking:
                /*
@@ -973,7 +993,7 @@ retry_blocking:
                 */
                ret = bch2_bucket_alloc_set_trans(trans, ptrs, &wp->stripe, &devs,
                                        nr_replicas, nr_effective, have_cache,
-                                       flags, wp->data_type, reserve, cl);
+                                       flags, wp->data_type, watermark, cl);
                if (ret &&
                    !bch2_err_matches(ret, BCH_ERR_transaction_restart) &&
                    !bch2_err_matches(ret, BCH_ERR_insufficient_devices) &&
@@ -981,7 +1001,6 @@ retry_blocking:
                        cl = _cl;
                        goto retry_blocking;
                }
-
        }
 
        return ret;
@@ -996,7 +1015,7 @@ static int open_bucket_add_buckets(struct btree_trans *trans,
                        unsigned nr_replicas,
                        unsigned *nr_effective,
                        bool *have_cache,
-                       enum alloc_reserve reserve,
+                       enum bch_watermark watermark,
                        unsigned flags,
                        struct closure *cl)
 {
@@ -1006,7 +1025,7 @@ static int open_bucket_add_buckets(struct btree_trans *trans,
                ret = __open_bucket_add_buckets(trans, ptrs, wp,
                                devs_have, target, erasure_code,
                                nr_replicas, nr_effective, have_cache,
-                               reserve, flags, cl);
+                               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) ||
@@ -1019,10 +1038,23 @@ static int open_bucket_add_buckets(struct btree_trans *trans,
        ret = __open_bucket_add_buckets(trans, ptrs, wp,
                        devs_have, target, false,
                        nr_replicas, nr_effective, have_cache,
-                       reserve, flags, cl);
+                       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)
 {
@@ -1034,8 +1066,12 @@ static bool should_drop_bucket(struct open_bucket *ob, struct bch_fs *c,
                unsigned i;
 
                if (!drop && ob->ec) {
+                       unsigned nr_blocks;
+
                        mutex_lock(&ob->ec->lock);
-                       for (i = 0; i < ob->ec->new_stripe.key.v.nr_blocks; i++) {
+                       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;
 
@@ -1158,9 +1194,12 @@ 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) {
@@ -1179,19 +1218,14 @@ static bool try_decrease_writepoints(struct bch_fs *c, unsigned old_nr)
        hlist_del_rcu(&wp->node);
        mutex_unlock(&c->write_points_hash_lock);
 
-       bch2_writepoint_stop(c, NULL, false, 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_norelock(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)
 {
@@ -1249,6 +1283,30 @@ 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:
  */
@@ -1259,7 +1317,7 @@ int bch2_alloc_sectors_start_trans(struct btree_trans *trans,
                             struct bch_devs_list *devs_have,
                             unsigned nr_replicas,
                             unsigned nr_replicas_required,
-                            enum alloc_reserve reserve,
+                            enum bch_watermark watermark,
                             unsigned flags,
                             struct closure *cl,
                             struct write_point **wp_ret)
@@ -1273,6 +1331,9 @@ int bch2_alloc_sectors_start_trans(struct btree_trans *trans,
        int ret;
        int i;
 
+       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);
@@ -1292,15 +1353,24 @@ retry:
                ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
                                              target, erasure_code,
                                              nr_replicas, &nr_effective,
-                                             &have_cache, reserve,
+                                             &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;
+               if (bch2_err_matches(ret, BCH_ERR_open_buckets_empty)) {
+                       int ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
+                                             target, erasure_code,
+                                             nr_replicas, &nr_effective,
+                                             &have_cache, watermark,
+                                             flags, cl);
+                       if (!ret ||
+                           bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
+                           bch2_err_matches(ret, BCH_ERR_open_buckets_empty))
+                               goto alloc_done;
+               }
 
                /*
                 * Only try to allocate cache (durability = 0 devices) from the
@@ -1311,14 +1381,13 @@ retry:
                ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
                                              0, erasure_code,
                                              nr_replicas, &nr_effective,
-                                             &have_cache, reserve,
+                                             &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, reserve,
+                                             &have_cache, watermark,
                                              flags, cl);
        }
 alloc_done:
@@ -1334,9 +1403,12 @@ 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;
 
@@ -1353,13 +1425,13 @@ 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) ||
@@ -1506,25 +1578,47 @@ static const char * const bch2_write_point_states[] = {
        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;
-       unsigned i;
 
+       prt_str(out, "Foreground write points\n");
        for (wp = c->write_points;
             wp < c->write_points + ARRAY_SIZE(c->write_points);
-            wp++) {
-               prt_printf(out, "%lu: ", wp->write_point);
-               prt_human_readable_u64(out, wp->sectors_allocated);
+            wp++)
+               bch2_write_point_to_text(out, c, wp);
 
-               prt_printf(out, " last wrote: ");
-               bch2_pr_time_units(out, sched_clock() - wp->last_used);
+       prt_str(out, "Copygc write point\n");
+       bch2_write_point_to_text(out, c, &c->copygc_write_point);
 
-               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_str(out, "Rebalance write point\n");
+       bch2_write_point_to_text(out, c, &c->rebalance_write_point);
 
-               prt_newline(out);
-       }
+       prt_str(out, "Btree write point\n");
+       bch2_write_point_to_text(out, c, &c->btree_write_point);
 }