X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libbcachefs%2Falloc_foreground.c;h=a961df74d421707468ed0868932d25a59fc218fb;hb=89b361f24a433a4a55e0032eca4e43045ea3f0d0;hp=fcb7311b184463313b965b01ef1785c1bdc09c43;hpb=5d507f795b0b679a67e972a48cbd0854c4ad0f02;p=bcachefs-tools-debian

diff --git a/libbcachefs/alloc_foreground.c b/libbcachefs/alloc_foreground.c
index fcb7311..a961df7 100644
--- a/libbcachefs/alloc_foreground.c
+++ b/libbcachefs/alloc_foreground.c
@@ -25,7 +25,7 @@
 #include "disk_groups.h"
 #include "ec.h"
 #include "error.h"
-#include "io.h"
+#include "io_write.h"
 #include "journal.h"
 #include "movinggc.h"
 #include "nocow_locking.h"
@@ -399,12 +399,23 @@ bch2_bucket_alloc_early(struct btree_trans *trans,
 			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) {
@@ -419,25 +430,38 @@ 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, watermark, a, s, cl);
+next:
+		citer.path->preserve = false;
+		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;
 	}
 
@@ -502,9 +526,14 @@ 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,
@@ -597,7 +626,7 @@ struct open_bucket *bch2_bucket_alloc(struct bch_fs *c, struct bch_dev *ca,
 	struct open_bucket *ob;
 
 	bch2_trans_do(c, NULL, NULL, 0,
-		      PTR_ERR_OR_ZERO(ob = bch2_bucket_alloc_trans(&trans, ca, watermark,
+		      PTR_ERR_OR_ZERO(ob = bch2_bucket_alloc_trans(trans, ca, watermark,
 							cl, &usage)));
 	return ob;
 }
@@ -668,11 +697,9 @@ 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);
@@ -775,7 +802,6 @@ 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;
 
@@ -805,8 +831,6 @@ 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);
@@ -989,7 +1013,6 @@ retry_blocking:
 			cl = _cl;
 			goto retry_blocking;
 		}
-
 	}
 
 	return ret;
@@ -1031,6 +1054,19 @@ static int open_bucket_add_buckets(struct btree_trans *trans,
 	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)
 {
@@ -1042,8 +1078,12 @@ 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;
 
@@ -1255,6 +1295,30 @@ 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:
  */
@@ -1279,6 +1343,9 @@ 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);
@@ -1340,6 +1407,9 @@ 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, ob);
@@ -1512,25 +1582,47 @@ 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);
 }