#include "journal.h"
#include "keylist.h"
#include "move.h"
+#include "nocow_locking.h"
#include "rebalance.h"
#include "subvolume.h"
#include "super.h"
#include "super-io.h"
#include <linux/blkdev.h>
+#include <linux/prefetch.h>
#include <linux/random.h>
#include <linux/sched/mm.h>
return blk_status_to_str(status);
}
+#ifndef CONFIG_BCACHEFS_NO_LATENCY_ACCT
+
static bool bch2_target_congested(struct bch_fs *c, u16 target)
{
const struct bch_devs_mask *devs;
__bch2_time_stats_update(&ca->io_latency[rw], submit_time, now);
}
+#else
+
+static bool bch2_target_congested(struct bch_fs *c, u16 target)
+{
+ return false;
+}
+
+#endif
+
/* Allocate, free from mempool: */
void bch2_bio_free_pages_pool(struct bch_fs *c, struct bio *bio)
struct btree_iter iter;
struct bkey_i *k;
struct bkey_i_inode_v3 *inode;
+ unsigned inode_update_flags = BTREE_UPDATE_NOJOURNAL;
int ret;
bch2_trans_iter_init(trans, &iter, BTREE_ID_inodes,
inode = bkey_i_to_inode_v3(k);
if (!(le64_to_cpu(inode->v.bi_flags) & BCH_INODE_I_SIZE_DIRTY) &&
- new_i_size > le64_to_cpu(inode->v.bi_size))
+ new_i_size > le64_to_cpu(inode->v.bi_size)) {
inode->v.bi_size = cpu_to_le64(new_i_size);
+ inode_update_flags = 0;
+ }
- le64_add_cpu(&inode->v.bi_sectors, i_sectors_delta);
+ if (i_sectors_delta) {
+ le64_add_cpu(&inode->v.bi_sectors, i_sectors_delta);
+ inode_update_flags = 0;
+ }
- inode->k.p.snapshot = iter.snapshot;
+ if (inode->k.p.snapshot != iter.snapshot) {
+ inode->k.p.snapshot = iter.snapshot;
+ inode_update_flags = 0;
+ }
ret = bch2_trans_update(trans, &iter, &inode->k_i,
- BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
+ BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
+ inode_update_flags);
err:
bch2_trans_iter_exit(trans, &iter);
return ret;
opts.data_replicas,
opts.data_replicas,
RESERVE_none, 0, &cl, &wp);
- if (ret == -EAGAIN) {
+ if (bch2_err_matches(ret, BCH_ERR_operation_blocked)) {
bch2_trans_unlock(trans);
closure_sync(&cl);
goto retry;
if (ret)
continue;
- BUG_ON(bkey_ge(iter->pos, end_pos));
-
bkey_init(&delete.k);
delete.k.p = iter->pos;
bch2_disk_reservation_put(c, &disk_res);
}
- BUG_ON(bkey_gt(iter->pos, end_pos));
-
return ret ?: ret2;
}
bio_sectors(&n->bio));
bio_set_dev(&n->bio, ca->disk_sb.bdev);
+
+ if (type != BCH_DATA_btree && unlikely(c->opts.no_data_io)) {
+ bio_endio(&n->bio);
+ continue;
+ }
+
submit_bio(&n->bio);
} else {
n->bio.bi_status = BLK_STS_REMOVED;
struct bch_fs *c = op->c;
bch2_disk_reservation_put(c, &op->res);
- percpu_ref_put(&c->writes);
+ bch2_write_ref_put(c, BCH_WRITE_REF_write);
bch2_keylist_free(&op->insert_keys, op->inline_keys);
bch2_time_stats_update(&c->times[BCH_TIME_data_write], op->start_time);
+ EBUG_ON(cl->parent);
closure_debug_destroy(cl);
if (op->end_io)
op->end_io(op);
op->written += sectors_start - keylist_sectors(keys);
- if (ret) {
+ if (ret && !bch2_err_matches(ret, EROFS)) {
struct bkey_i *k = bch2_keylist_front(&op->insert_keys);
bch_err_inum_offset_ratelimited(c,
k->k.p.inode, k->k.p.offset << 9,
"write error while doing btree update: %s",
bch2_err_str(ret));
- goto err;
}
+
+ if (ret)
+ goto err;
}
out:
/* If some a bucket wasn't written, we can't erasure code it: */
goto out;
}
+static inline void __wp_update_state(struct write_point *wp, enum write_point_state state)
+{
+ if (state != wp->state) {
+ u64 now = ktime_get_ns();
+
+ if (wp->last_state_change &&
+ time_after64(now, wp->last_state_change))
+ wp->time[wp->state] += now - wp->last_state_change;
+ wp->state = state;
+ wp->last_state_change = now;
+ }
+}
+
+static inline void wp_update_state(struct write_point *wp, bool running)
+{
+ enum write_point_state state;
+
+ state = running ? WRITE_POINT_running :
+ !list_empty(&wp->writes) ? WRITE_POINT_waiting_io
+ : WRITE_POINT_stopped;
+
+ __wp_update_state(wp, state);
+}
+
static void bch2_write_index(struct closure *cl)
{
struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
struct workqueue_struct *wq = index_update_wq(op);
barrier();
+
+ /*
+ * We're not using wp->writes_lock here, so this is racey: that's ok,
+ * because this is just for diagnostic purposes, and we're running out
+ * of interrupt context here so if we were to take the log we'd have to
+ * switch to spin_lock_irq()/irqsave(), which is not free:
+ */
+ if (wp->state == WRITE_POINT_waiting_io)
+ __wp_update_state(wp, WRITE_POINT_waiting_work);
+
op->btree_update_ready = true;
queue_work(wq, &wp->index_update_work);
}
+static inline void bch2_write_queue(struct bch_write_op *op, struct write_point *wp)
+{
+ op->btree_update_ready = false;
+ op->wp = wp;
+
+ spin_lock(&wp->writes_lock);
+ list_add_tail(&op->wp_list, &wp->writes);
+ if (wp->state == WRITE_POINT_stopped)
+ __wp_update_state(wp, WRITE_POINT_waiting_io);
+ spin_unlock(&wp->writes_lock);
+}
+
void bch2_write_point_do_index_updates(struct work_struct *work)
{
struct write_point *wp =
while (1) {
spin_lock(&wp->writes_lock);
- op = list_first_entry_or_null(&wp->writes, struct bch_write_op, wp_list);
- if (op && !op->btree_update_ready)
- op = NULL;
- if (op)
- list_del(&op->wp_list);
+ list_for_each_entry(op, &wp->writes, wp_list)
+ if (op->btree_update_ready) {
+ list_del(&op->wp_list);
+ goto unlock;
+ }
+ op = NULL;
+unlock:
+ wp_update_state(wp, op != NULL);
spin_unlock(&wp->writes_lock);
if (!op)
break;
+ op->flags |= BCH_WRITE_IN_WORKER;
+
__bch2_write_index(op);
if (!(op->flags & BCH_WRITE_DONE))
if (wbio->put_bio)
bio_put(bio);
- if (parent) {
+ if (parent)
bio_endio(&parent->bio);
- return;
- }
-
- closure_put(cl);
+ else
+ closure_put(cl);
}
static void init_append_extent(struct bch_write_op *op,
struct bversion version,
struct bch_extent_crc_unpacked crc)
{
- struct bch_fs *c = op->c;
struct bkey_i_extent *e;
op->pos.offset += crc.uncompressed_size;
crc.nonce)
bch2_extent_crc_append(&e->k_i, crc);
- bch2_alloc_sectors_append_ptrs_inlined(c, wp, &e->k_i, crc.compressed_size,
+ bch2_alloc_sectors_append_ptrs_inlined(op->c, wp, &e->k_i, crc.compressed_size,
op->flags & BCH_WRITE_CACHED);
bch2_keylist_push(&op->insert_keys);
bkey_start_pos(&orig->k), orig->k.p,
BTREE_ITER_INTENT, k,
NULL, NULL, BTREE_INSERT_NOFAIL, ({
- BUG_ON(bkey_ge(bkey_start_pos(k.k), orig->k.p));
-
bch2_nocow_write_convert_one_unwritten(&trans, &iter, orig, k, op->new_i_size);
}));
- if (ret) {
+ if (ret && !bch2_err_matches(ret, EROFS)) {
struct bkey_i *k = bch2_keylist_front(&op->insert_keys);
bch_err_inum_offset_ratelimited(c,
k->k.p.inode, k->k.p.offset << 9,
"write error while doing btree update: %s",
bch2_err_str(ret));
+ }
+
+ if (ret) {
op->error = ret;
break;
}
struct {
struct bpos b;
unsigned gen;
- two_state_lock_t *l;
+ struct nocow_lock_bucket *l;
} buckets[BCH_REPLICAS_MAX];
unsigned nr_buckets = 0;
u32 snapshot;
buckets[nr_buckets].b = PTR_BUCKET_POS(c, ptr);
buckets[nr_buckets].gen = ptr->gen;
buckets[nr_buckets].l =
- bucket_nocow_lock(&c->nocow_locks, buckets[nr_buckets].b);
+ bucket_nocow_lock(&c->nocow_locks,
+ bucket_to_u64(buckets[nr_buckets].b));
prefetch(buckets[nr_buckets].l);
nr_buckets++;
for (i = 0; i < nr_buckets; i++) {
struct bch_dev *ca = bch_dev_bkey_exists(c, buckets[i].b.inode);
- two_state_lock_t *l = buckets[i].l;
+ struct nocow_lock_bucket *l = buckets[i].l;
bool stale;
- if (!bch2_two_state_trylock(l, BUCKET_NOCOW_LOCK_UPDATE))
- __bch2_bucket_nocow_lock(&c->nocow_locks, l, BUCKET_NOCOW_LOCK_UPDATE);
+ __bch2_bucket_nocow_lock(&c->nocow_locks, l,
+ bucket_to_u64(buckets[i].b),
+ BUCKET_NOCOW_LOCK_UPDATE);
rcu_read_lock();
stale = gen_after(*bucket_gen(ca, buckets[i].b.offset), buckets[i].gen);
nofs_flags = memalloc_nofs_save();
- if (unlikely(op->opts.nocow)) {
+ if (unlikely(op->opts.nocow && c->opts.nocow_enabled)) {
bch2_nocow_write(op);
if (op->flags & BCH_WRITE_DONE)
goto out_nofs_restore;
BCH_WRITE_ONLY_SPECIFIED_DEVS))
? NULL : &op->cl, &wp));
if (unlikely(ret)) {
- if (ret == -EAGAIN)
+ if (bch2_err_matches(ret, BCH_ERR_operation_blocked))
break;
goto err;
}
+ EBUG_ON(!wp);
+
+ bch2_open_bucket_get(c, wp, &op->open_buckets);
ret = bch2_write_extent(op, wp, &bio);
- if (ret >= 0)
- bch2_open_bucket_get(c, wp, &op->open_buckets);
bch2_alloc_sectors_done_inlined(c, wp);
err:
if (ret <= 0) {
- if (!(op->flags & BCH_WRITE_SYNC)) {
- spin_lock(&wp->writes_lock);
- op->wp = wp;
- list_add_tail(&op->wp_list, &wp->writes);
- spin_unlock(&wp->writes_lock);
- }
-
op->flags |= BCH_WRITE_DONE;
if (ret < 0) {
* synchronously here if we weren't able to submit all of the IO at
* once, as that signals backpressure to the caller.
*/
- if ((op->flags & BCH_WRITE_SYNC) || !(op->flags & BCH_WRITE_DONE)) {
+ if ((op->flags & BCH_WRITE_SYNC) ||
+ (!(op->flags & BCH_WRITE_DONE) &&
+ !(op->flags & BCH_WRITE_IN_WORKER))) {
closure_sync(&op->cl);
__bch2_write_index(op);
goto again;
bch2_write_done(&op->cl);
} else {
+ bch2_write_queue(op, wp);
continue_at(&op->cl, bch2_write_index, NULL);
}
out_nofs_restore:
unsigned sectors;
int ret;
+ op->flags |= BCH_WRITE_WROTE_DATA_INLINE;
+ op->flags |= BCH_WRITE_DONE;
+
bch2_check_set_feature(op->c, BCH_FEATURE_inline_data);
ret = bch2_keylist_realloc(&op->insert_keys, op->inline_keys,
set_bkey_val_bytes(&id->k, data_len);
bch2_keylist_push(&op->insert_keys);
- op->flags |= BCH_WRITE_WROTE_DATA_INLINE;
- op->flags |= BCH_WRITE_DONE;
-
__bch2_write_index(op);
err:
bch2_write_done(&op->cl);
}
if (c->opts.nochanges ||
- !percpu_ref_tryget_live(&c->writes)) {
- op->error = -EROFS;
+ !bch2_write_ref_tryget(c, BCH_WRITE_REF_write)) {
+ op->error = -BCH_ERR_erofs_no_writes;
goto err;
}
{
int ret;
+ bch2_data_update_exit(&op->write);
+
ret = rhashtable_remove_fast(&c->promote_table, &op->hash,
bch_promote_params);
BUG_ON(ret);
- percpu_ref_put(&c->writes);
+ bch2_write_ref_put(c, BCH_WRITE_REF_promote);
kfree_rcu(op, rcu);
}
bch2_time_stats_update(&c->times[BCH_TIME_data_promote],
op->start_time);
-
- bch2_data_update_exit(&op->write);
promote_free(c, op);
}
bch2_data_update_read_done(&op->write, rbio->pick.crc);
}
-static struct promote_op *__promote_alloc(struct bch_fs *c,
+static struct promote_op *__promote_alloc(struct btree_trans *trans,
enum btree_id btree_id,
struct bkey_s_c k,
struct bpos pos,
unsigned sectors,
struct bch_read_bio **rbio)
{
+ struct bch_fs *c = trans->c;
struct promote_op *op = NULL;
struct bio *bio;
unsigned pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
int ret;
- if (!percpu_ref_tryget_live(&c->writes))
+ if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_promote))
return NULL;
op = kzalloc(sizeof(*op) + sizeof(struct bio_vec) * pages, GFP_NOIO);
bio = &op->write.op.wbio.bio;
bio_init(bio, NULL, bio->bi_inline_vecs, pages, 0);
- ret = bch2_data_update_init(c, &op->write,
+ ret = bch2_data_update_init(trans, NULL, &op->write,
writepoint_hashed((unsigned long) current),
opts,
(struct data_update_opts) {
.write_flags = BCH_WRITE_ALLOC_NOWAIT|BCH_WRITE_CACHED,
},
btree_id, k);
+ if (ret == -BCH_ERR_nocow_lock_blocked) {
+ ret = rhashtable_remove_fast(&c->promote_table, &op->hash,
+ bch_promote_params);
+ BUG_ON(ret);
+ goto err;
+ }
+
BUG_ON(ret);
op->write.op.end_io = promote_done;
kfree(*rbio);
*rbio = NULL;
kfree(op);
- percpu_ref_put(&c->writes);
+ bch2_write_ref_put(c, BCH_WRITE_REF_promote);
return NULL;
}
noinline
-static struct promote_op *promote_alloc(struct bch_fs *c,
- struct bvec_iter iter,
- struct bkey_s_c k,
- struct extent_ptr_decoded *pick,
- struct bch_io_opts opts,
- unsigned flags,
- struct bch_read_bio **rbio,
- bool *bounce,
- bool *read_full)
+static struct promote_op *promote_alloc(struct btree_trans *trans,
+ struct bvec_iter iter,
+ struct bkey_s_c k,
+ struct extent_ptr_decoded *pick,
+ struct bch_io_opts opts,
+ unsigned flags,
+ struct bch_read_bio **rbio,
+ bool *bounce,
+ bool *read_full)
{
+ struct bch_fs *c = trans->c;
bool promote_full = *read_full || READ_ONCE(c->promote_whole_extents);
/* data might have to be decompressed in the write path: */
unsigned sectors = promote_full
if (!should_promote(c, k, pos, opts, flags))
return NULL;
- promote = __promote_alloc(c,
+ promote = __promote_alloc(trans,
k.k->type == KEY_TYPE_reflink_v
? BTREE_ID_reflink
: BTREE_ID_extents,
}
csum = bch2_checksum_bio(c, crc.csum_type, nonce, src);
- if (bch2_crc_cmp(csum, rbio->pick.crc.csum))
+ if (bch2_crc_cmp(csum, rbio->pick.crc.csum) && !c->opts.no_data_io)
goto csum_err;
/*
}
if (orig->opts.promote_target)
- promote = promote_alloc(c, iter, k, &pick, orig->opts, flags,
+ promote = promote_alloc(trans, iter, k, &pick, orig->opts, flags,
&rbio, &bounce, &read_full);
if (!read_full) {
bio_sectors(&rbio->bio));
bio_set_dev(&rbio->bio, ca->disk_sb.bdev);
- if (likely(!(flags & BCH_READ_IN_RETRY)))
- submit_bio(&rbio->bio);
- else
- submit_bio_wait(&rbio->bio);
+ if (unlikely(c->opts.no_data_io)) {
+ if (likely(!(flags & BCH_READ_IN_RETRY)))
+ bio_endio(&rbio->bio);
+ } else {
+ if (likely(!(flags & BCH_READ_IN_RETRY)))
+ submit_bio(&rbio->bio);
+ else
+ submit_bio_wait(&rbio->bio);
+ }
+
+ /*
+ * We just submitted IO which may block, we expect relock fail
+ * events and shouldn't count them:
+ */
+ trans->notrace_relock_fail = true;
} else {
/* Attempting reconstruct read: */
if (bch2_ec_read_extent(c, rbio)) {
int bch2_fs_io_init(struct bch_fs *c)
{
- unsigned i;
-
- for (i = 0; i < ARRAY_SIZE(c->nocow_locks.l); i++)
- two_state_lock_init(&c->nocow_locks.l[i]);
-
if (bioset_init(&c->bio_read, 1, offsetof(struct bch_read_bio, bio),
BIOSET_NEED_BVECS) ||
bioset_init(&c->bio_read_split, 1, offsetof(struct bch_read_bio, bio),
projects / bcachefs-tools-debian / blobdiff