+// SPDX-License-Identifier: GPL-2.0
/*
* Some low level IO code, and hacks for various block layer limitations
*
*/
#include "bcachefs.h"
-#include "alloc.h"
+#include "alloc_background.h"
+#include "alloc_foreground.h"
+#include "bkey_buf.h"
#include "bset.h"
#include "btree_update.h"
#include "buckets.h"
#include "compress.h"
#include "clock.h"
#include "debug.h"
+#include "disk_groups.h"
+#include "ec.h"
#include "error.h"
-#include "extents.h"
+#include "extent_update.h"
+#include "inode.h"
#include "io.h"
#include "journal.h"
#include "keylist.h"
#include "move.h"
+#include "rebalance.h"
+#include "subvolume.h"
+#include "super.h"
#include "super-io.h"
#include <linux/blkdev.h>
#include <linux/random.h>
+#include <linux/sched/mm.h>
#include <trace/events/bcachefs.h>
-static inline void __bio_inc_remaining(struct bio *bio)
+const char *bch2_blk_status_to_str(blk_status_t status)
{
- bio_set_flag(bio, BIO_CHAIN);
- smp_mb__before_atomic();
- atomic_inc(&bio->__bi_remaining);
+ if (status == BLK_STS_REMOVED)
+ return "device removed";
+ return blk_status_to_str(status);
+}
+
+static bool bch2_target_congested(struct bch_fs *c, u16 target)
+{
+ const struct bch_devs_mask *devs;
+ unsigned d, nr = 0, total = 0;
+ u64 now = local_clock(), last;
+ s64 congested;
+ struct bch_dev *ca;
+
+ if (!target)
+ return false;
+
+ rcu_read_lock();
+ devs = bch2_target_to_mask(c, target) ?:
+ &c->rw_devs[BCH_DATA_user];
+
+ for_each_set_bit(d, devs->d, BCH_SB_MEMBERS_MAX) {
+ ca = rcu_dereference(c->devs[d]);
+ if (!ca)
+ continue;
+
+ congested = atomic_read(&ca->congested);
+ last = READ_ONCE(ca->congested_last);
+ if (time_after64(now, last))
+ congested -= (now - last) >> 12;
+
+ total += max(congested, 0LL);
+ nr++;
+ }
+ rcu_read_unlock();
+
+ return bch2_rand_range(nr * CONGESTED_MAX) < total;
+}
+
+static inline void bch2_congested_acct(struct bch_dev *ca, u64 io_latency,
+ u64 now, int rw)
+{
+ u64 latency_capable =
+ ca->io_latency[rw].quantiles.entries[QUANTILE_IDX(1)].m;
+ /* ideally we'd be taking into account the device's variance here: */
+ u64 latency_threshold = latency_capable << (rw == READ ? 2 : 3);
+ s64 latency_over = io_latency - latency_threshold;
+
+ if (latency_threshold && latency_over > 0) {
+ /*
+ * bump up congested by approximately latency_over * 4 /
+ * latency_threshold - we don't need much accuracy here so don't
+ * bother with the divide:
+ */
+ if (atomic_read(&ca->congested) < CONGESTED_MAX)
+ atomic_add(latency_over >>
+ max_t(int, ilog2(latency_threshold) - 2, 0),
+ &ca->congested);
+
+ ca->congested_last = now;
+ } else if (atomic_read(&ca->congested) > 0) {
+ atomic_dec(&ca->congested);
+ }
+}
+
+void bch2_latency_acct(struct bch_dev *ca, u64 submit_time, int rw)
+{
+ atomic64_t *latency = &ca->cur_latency[rw];
+ u64 now = local_clock();
+ u64 io_latency = time_after64(now, submit_time)
+ ? now - submit_time
+ : 0;
+ u64 old, new, v = atomic64_read(latency);
+
+ do {
+ old = v;
+
+ /*
+ * If the io latency was reasonably close to the current
+ * latency, skip doing the update and atomic operation - most of
+ * the time:
+ */
+ if (abs((int) (old - io_latency)) < (old >> 1) &&
+ now & ~(~0U << 5))
+ break;
+
+ new = ewma_add(old, io_latency, 5);
+ } while ((v = atomic64_cmpxchg(latency, old, new)) != old);
+
+ bch2_congested_acct(ca, io_latency, now, rw);
+
+ __bch2_time_stats_update(&ca->io_latency[rw], submit_time, now);
}
/* Allocate, free from mempool: */
void bch2_bio_free_pages_pool(struct bch_fs *c, struct bio *bio)
{
+ struct bvec_iter_all iter;
struct bio_vec *bv;
- unsigned i;
- bio_for_each_segment_all(bv, bio, i)
+ bio_for_each_segment_all(bv, bio, iter)
if (bv->bv_page != ZERO_PAGE(0))
mempool_free(bv->bv_page, &c->bio_bounce_pages);
bio->bi_vcnt = 0;
}
-static void bch2_bio_alloc_page_pool(struct bch_fs *c, struct bio *bio,
- bool *using_mempool)
+static struct page *__bio_alloc_page_pool(struct bch_fs *c, bool *using_mempool)
{
- struct bio_vec *bv = &bio->bi_io_vec[bio->bi_vcnt++];
+ struct page *page;
if (likely(!*using_mempool)) {
- bv->bv_page = alloc_page(GFP_NOIO);
- if (unlikely(!bv->bv_page)) {
+ page = alloc_page(GFP_NOIO);
+ if (unlikely(!page)) {
mutex_lock(&c->bio_bounce_pages_lock);
*using_mempool = true;
goto pool_alloc;
}
} else {
pool_alloc:
- bv->bv_page = mempool_alloc(&c->bio_bounce_pages, GFP_NOIO);
+ page = mempool_alloc(&c->bio_bounce_pages, GFP_NOIO);
}
- bv->bv_len = PAGE_SIZE;
- bv->bv_offset = 0;
+ return page;
}
void bch2_bio_alloc_pages_pool(struct bch_fs *c, struct bio *bio,
- size_t bytes)
+ size_t size)
{
bool using_mempool = false;
- bio->bi_iter.bi_size = bytes;
+ while (size) {
+ struct page *page = __bio_alloc_page_pool(c, &using_mempool);
+ unsigned len = min_t(size_t, PAGE_SIZE, size);
- while (bio->bi_vcnt < DIV_ROUND_UP(bytes, PAGE_SIZE))
- bch2_bio_alloc_page_pool(c, bio, &using_mempool);
+ BUG_ON(!bio_add_page(bio, page, len, 0));
+ size -= len;
+ }
if (using_mempool)
mutex_unlock(&c->bio_bounce_pages_lock);
}
-/* Bios with headers */
+/* Extent update path: */
+
+int bch2_sum_sector_overwrites(struct btree_trans *trans,
+ struct btree_iter *extent_iter,
+ struct bkey_i *new,
+ bool *usage_increasing,
+ s64 *i_sectors_delta,
+ s64 *disk_sectors_delta)
+{
+ struct bch_fs *c = trans->c;
+ struct btree_iter iter;
+ struct bkey_s_c old;
+ unsigned new_replicas = bch2_bkey_replicas(c, bkey_i_to_s_c(new));
+ bool new_compressed = bch2_bkey_sectors_compressed(bkey_i_to_s_c(new));
+ int ret = 0;
+
+ *usage_increasing = false;
+ *i_sectors_delta = 0;
+ *disk_sectors_delta = 0;
+
+ bch2_trans_copy_iter(&iter, extent_iter);
+
+ for_each_btree_key_continue_norestart(iter, BTREE_ITER_SLOTS, old, ret) {
+ s64 sectors = min(new->k.p.offset, old.k->p.offset) -
+ max(bkey_start_offset(&new->k),
+ bkey_start_offset(old.k));
+
+ *i_sectors_delta += sectors *
+ (bkey_extent_is_allocation(&new->k) -
+ bkey_extent_is_allocation(old.k));
+
+ *disk_sectors_delta += sectors * bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(new));
+ *disk_sectors_delta -= new->k.p.snapshot == old.k->p.snapshot
+ ? sectors * bch2_bkey_nr_ptrs_fully_allocated(old)
+ : 0;
+
+ if (!*usage_increasing &&
+ (new->k.p.snapshot != old.k->p.snapshot ||
+ new_replicas > bch2_bkey_replicas(c, old) ||
+ (!new_compressed && bch2_bkey_sectors_compressed(old))))
+ *usage_increasing = true;
+
+ if (bkey_cmp(old.k->p, new->k.p) >= 0)
+ break;
+ }
+
+ bch2_trans_iter_exit(trans, &iter);
+ return ret;
+}
+
+int bch2_extent_update(struct btree_trans *trans,
+ subvol_inum inum,
+ struct btree_iter *iter,
+ struct bkey_i *k,
+ struct disk_reservation *disk_res,
+ u64 *journal_seq,
+ u64 new_i_size,
+ s64 *i_sectors_delta_total,
+ bool check_enospc)
+{
+ struct btree_iter inode_iter;
+ struct bch_inode_unpacked inode_u;
+ struct bpos next_pos;
+ bool usage_increasing;
+ s64 i_sectors_delta = 0, disk_sectors_delta = 0;
+ int ret;
+
+ /*
+ * This traverses us the iterator without changing iter->path->pos to
+ * search_key() (which is pos + 1 for extents): we want there to be a
+ * path already traversed at iter->pos because
+ * bch2_trans_extent_update() will use it to attempt extent merging
+ */
+ ret = __bch2_btree_iter_traverse(iter);
+ if (ret)
+ return ret;
+
+ ret = bch2_extent_trim_atomic(trans, iter, k);
+ if (ret)
+ return ret;
+
+ new_i_size = min(k->k.p.offset << 9, new_i_size);
+ next_pos = k->k.p;
+
+ ret = bch2_sum_sector_overwrites(trans, iter, k,
+ &usage_increasing,
+ &i_sectors_delta,
+ &disk_sectors_delta);
+ if (ret)
+ return ret;
+
+ if (disk_res &&
+ disk_sectors_delta > (s64) disk_res->sectors) {
+ ret = bch2_disk_reservation_add(trans->c, disk_res,
+ disk_sectors_delta - disk_res->sectors,
+ !check_enospc || !usage_increasing
+ ? BCH_DISK_RESERVATION_NOFAIL : 0);
+ if (ret)
+ return ret;
+ }
+
+ ret = bch2_inode_peek(trans, &inode_iter, &inode_u, inum,
+ BTREE_ITER_INTENT);
+ if (ret)
+ return ret;
+
+ if (!(inode_u.bi_flags & BCH_INODE_I_SIZE_DIRTY) &&
+ new_i_size > inode_u.bi_size)
+ inode_u.bi_size = new_i_size;
+
+ inode_u.bi_sectors += i_sectors_delta;
+
+ ret = bch2_trans_update(trans, iter, k, 0) ?:
+ bch2_inode_write(trans, &inode_iter, &inode_u) ?:
+ bch2_trans_commit(trans, disk_res, journal_seq,
+ BTREE_INSERT_NOCHECK_RW|
+ BTREE_INSERT_NOFAIL);
+ bch2_trans_iter_exit(trans, &inode_iter);
+
+ if (ret)
+ return ret;
+
+ if (i_sectors_delta_total)
+ *i_sectors_delta_total += i_sectors_delta;
+ bch2_btree_iter_set_pos(iter, next_pos);
+
+ return 0;
+}
+
+/*
+ * Returns -EINTR if we had to drop locks:
+ */
+int bch2_fpunch_at(struct btree_trans *trans, struct btree_iter *iter,
+ subvol_inum inum, u64 end,
+ s64 *i_sectors_delta)
+{
+ struct bch_fs *c = trans->c;
+ unsigned max_sectors = KEY_SIZE_MAX & (~0 << c->block_bits);
+ struct bpos end_pos = POS(inum.inum, end);
+ struct bkey_s_c k;
+ int ret = 0, ret2 = 0;
+ u32 snapshot;
+
+ while (!ret || ret == -EINTR) {
+ struct disk_reservation disk_res =
+ bch2_disk_reservation_init(c, 0);
+ struct bkey_i delete;
+
+ if (ret)
+ ret2 = ret;
+
+ bch2_trans_begin(trans);
+
+ ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
+ if (ret)
+ continue;
+
+ bch2_btree_iter_set_snapshot(iter, snapshot);
+
+ k = bch2_btree_iter_peek(iter);
+ if (bkey_cmp(iter->pos, end_pos) >= 0) {
+ bch2_btree_iter_set_pos(iter, end_pos);
+ break;
+ }
+
+ ret = bkey_err(k);
+ if (ret)
+ continue;
+
+ bkey_init(&delete.k);
+ delete.k.p = iter->pos;
+
+ /* create the biggest key we can */
+ bch2_key_resize(&delete.k, max_sectors);
+ bch2_cut_back(end_pos, &delete);
+
+ ret = bch2_extent_update(trans, inum, iter, &delete,
+ &disk_res, NULL,
+ 0, i_sectors_delta, false);
+ bch2_disk_reservation_put(c, &disk_res);
+ }
+
+ return ret ?: ret2;
+}
+
+int bch2_fpunch(struct bch_fs *c, subvol_inum inum, u64 start, u64 end,
+ s64 *i_sectors_delta)
+{
+ struct btree_trans trans;
+ struct btree_iter iter;
+ int ret;
+
+ bch2_trans_init(&trans, c, BTREE_ITER_MAX, 1024);
+ bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents,
+ POS(inum.inum, start),
+ BTREE_ITER_INTENT);
+
+ ret = bch2_fpunch_at(&trans, &iter, inum, end, i_sectors_delta);
+
+ bch2_trans_iter_exit(&trans, &iter);
+ bch2_trans_exit(&trans);
+
+ return ret == -EINTR ? 0 : ret;
+}
+
+int bch2_write_index_default(struct bch_write_op *op)
+{
+ struct bch_fs *c = op->c;
+ struct bkey_buf sk;
+ struct open_bucket *ec_ob = ec_open_bucket(c, &op->open_buckets);
+ struct keylist *keys = &op->insert_keys;
+ struct bkey_i *k = bch2_keylist_front(keys);
+ struct btree_trans trans;
+ struct btree_iter iter;
+ subvol_inum inum = {
+ .subvol = op->subvol,
+ .inum = k->k.p.inode,
+ };
+ int ret;
+
+ BUG_ON(!inum.subvol);
+
+ bch2_bkey_buf_init(&sk);
+ bch2_trans_init(&trans, c, BTREE_ITER_MAX, 1024);
+
+ do {
+ bch2_trans_begin(&trans);
+
+ k = bch2_keylist_front(keys);
+ bch2_bkey_buf_copy(&sk, c, k);
+
+ ret = bch2_subvolume_get_snapshot(&trans, inum.subvol,
+ &sk.k->k.p.snapshot);
+ if (ret == -EINTR)
+ continue;
+ if (ret)
+ break;
+
+ bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents,
+ bkey_start_pos(&sk.k->k),
+ BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
+
+ ret = bch2_extent_update(&trans, inum, &iter, sk.k,
+ &op->res, op_journal_seq(op),
+ op->new_i_size, &op->i_sectors_delta,
+ op->flags & BCH_WRITE_CHECK_ENOSPC);
+ bch2_trans_iter_exit(&trans, &iter);
+
+ if (ret == -EINTR)
+ continue;
+ if (ret)
+ break;
+
+ if (ec_ob)
+ bch2_ob_add_backpointer(c, ec_ob, &sk.k->k);
+
+ if (bkey_cmp(iter.pos, k->k.p) >= 0)
+ bch2_keylist_pop_front(&op->insert_keys);
+ else
+ bch2_cut_front(iter.pos, k);
+ } while (!bch2_keylist_empty(keys));
+
+ bch2_trans_exit(&trans);
+ bch2_bkey_buf_exit(&sk, c);
+
+ return ret;
+}
+
+/* Writes */
void bch2_submit_wbio_replicas(struct bch_write_bio *wbio, struct bch_fs *c,
+ enum bch_data_type type,
const struct bkey_i *k)
{
- struct bkey_s_c_extent e = bkey_i_to_s_c_extent(k);
+ struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(bkey_i_to_s_c(k));
const struct bch_extent_ptr *ptr;
struct bch_write_bio *n;
struct bch_dev *ca;
BUG_ON(c->opts.nochanges);
- wbio->split = false;
- wbio->c = c;
+ bkey_for_each_ptr(ptrs, ptr) {
+ BUG_ON(ptr->dev >= BCH_SB_MEMBERS_MAX ||
+ !c->devs[ptr->dev]);
- extent_for_each_ptr(e, ptr) {
- ca = c->devs[ptr->dev];
+ ca = bch_dev_bkey_exists(c, ptr->dev);
- if (ptr + 1 < &extent_entry_last(e)->ptr) {
+ if (to_entry(ptr + 1) < ptrs.end) {
n = to_wbio(bio_clone_fast(&wbio->bio, GFP_NOIO,
&ca->replica_set));
n->bio.bi_end_io = wbio->bio.bi_end_io;
n->bio.bi_private = wbio->bio.bi_private;
- n->c = c;
- n->orig = &wbio->bio;
- n->bounce = false;
+ n->parent = wbio;
n->split = true;
+ n->bounce = false;
n->put_bio = true;
n->bio.bi_opf = wbio->bio.bi_opf;
- __bio_inc_remaining(n->orig);
+ bio_inc_remaining(&wbio->bio);
} else {
n = wbio;
+ n->split = false;
}
- if (!journal_flushes_device(ca))
- n->bio.bi_opf |= REQ_FUA;
-
- n->ca = ca;
- n->submit_time_us = local_clock_us();
+ n->c = c;
+ n->dev = ptr->dev;
+ n->have_ioref = bch2_dev_get_ioref(ca,
+ type == BCH_DATA_btree ? READ : WRITE);
+ n->submit_time = local_clock();
n->bio.bi_iter.bi_sector = ptr->offset;
- if (likely(percpu_ref_tryget(&ca->io_ref))) {
- n->have_io_ref = true;
- n->bio.bi_bdev = ca->disk_sb.bdev;
- generic_make_request(&n->bio);
+ if (likely(n->have_ioref)) {
+ this_cpu_add(ca->io_done->sectors[WRITE][type],
+ bio_sectors(&n->bio));
+
+ bio_set_dev(&n->bio, ca->disk_sb.bdev);
+ submit_bio(&n->bio);
} else {
- n->have_io_ref = false;
- bcache_io_error(c, &n->bio, "device has been removed");
+ n->bio.bi_status = BLK_STS_REMOVED;
bio_endio(&n->bio);
}
}
}
-/* IO errors */
-
-/* Writes */
-
-static struct workqueue_struct *index_update_wq(struct bch_write_op *op)
-{
- return op->alloc_reserve == RESERVE_MOVINGGC
- ? op->c->copygc_wq
- : op->c->wq;
-}
-
static void __bch2_write(struct closure *);
static void bch2_write_done(struct closure *cl)
{
struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
-
- BUG_ON(!(op->flags & BCH_WRITE_DONE));
+ struct bch_fs *c = op->c;
if (!op->error && (op->flags & BCH_WRITE_FLUSH))
- op->error = bch2_journal_error(&op->c->journal);
+ op->error = bch2_journal_error(&c->journal);
- bch2_disk_reservation_put(op->c, &op->res);
- percpu_ref_put(&op->c->writes);
+ bch2_disk_reservation_put(c, &op->res);
+ percpu_ref_put(&c->writes);
bch2_keylist_free(&op->insert_keys, op->inline_keys);
- closure_return(cl);
-}
-static u64 keylist_sectors(struct keylist *keys)
-{
- struct bkey_i *k;
- u64 ret = 0;
+ bch2_time_stats_update(&c->times[BCH_TIME_data_write], op->start_time);
- for_each_keylist_key(keys, k)
- ret += k->k.size;
-
- return ret;
+ if (op->end_io) {
+ EBUG_ON(cl->parent);
+ closure_debug_destroy(cl);
+ op->end_io(op);
+ } else {
+ closure_return(cl);
+ }
}
-static int bch2_write_index_default(struct bch_write_op *op)
+/**
+ * bch_write_index - after a write, update index to point to new data
+ */
+static void __bch2_write_index(struct bch_write_op *op)
{
+ struct bch_fs *c = op->c;
struct keylist *keys = &op->insert_keys;
- struct btree_iter iter;
+ struct bch_extent_ptr *ptr;
+ struct bkey_i *src, *dst = keys->keys, *n, *k;
+ unsigned dev;
int ret;
- bch2_btree_iter_init_intent(&iter, op->c, BTREE_ID_EXTENTS,
- bkey_start_pos(&bch2_keylist_front(keys)->k));
+ for (src = keys->keys; src != keys->top; src = n) {
+ n = bkey_next(src);
- ret = bch2_btree_insert_list_at(&iter, keys, &op->res,
- NULL, op_journal_seq(op),
- BTREE_INSERT_NOFAIL);
- bch2_btree_iter_unlock(&iter);
+ if (bkey_extent_is_direct_data(&src->k)) {
+ bch2_bkey_drop_ptrs(bkey_i_to_s(src), ptr,
+ test_bit(ptr->dev, op->failed.d));
- return ret;
-}
+ if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(src))) {
+ ret = -EIO;
+ goto err;
+ }
+ }
-/**
- * bch_write_index - after a write, update index to point to new data
- */
-static void bch2_write_index(struct closure *cl)
-{
- struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
- struct bch_fs *c = op->c;
- struct keylist *keys = &op->insert_keys;
- unsigned i;
+ if (dst != src)
+ memmove_u64s_down(dst, src, src->u64s);
+ dst = bkey_next(dst);
+ }
- op->flags |= BCH_WRITE_LOOPED;
+ keys->top = dst;
+
+ /*
+ * probably not the ideal place to hook this in, but I don't
+ * particularly want to plumb io_opts all the way through the btree
+ * update stack right now
+ */
+ for_each_keylist_key(keys, k) {
+ bch2_rebalance_add_key(c, bkey_i_to_s_c(k), &op->opts);
+
+ if (bch2_bkey_is_incompressible(bkey_i_to_s_c(k)))
+ bch2_check_set_feature(op->c, BCH_FEATURE_incompressible);
+
+ }
if (!bch2_keylist_empty(keys)) {
u64 sectors_start = keylist_sectors(keys);
int ret = op->index_update_fn(op);
+ BUG_ON(ret == -EINTR);
BUG_ON(keylist_sectors(keys) && !ret);
op->written += sectors_start - keylist_sectors(keys);
if (ret) {
- __bcache_io_error(c, "btree IO error %i", ret);
+ bch_err_inum_ratelimited(c, op->pos.inode,
+ "write error %i from btree update", ret);
op->error = ret;
}
}
+out:
+ /* If some a bucket wasn't written, we can't erasure code it: */
+ for_each_set_bit(dev, op->failed.d, BCH_SB_MEMBERS_MAX)
+ bch2_open_bucket_write_error(c, &op->open_buckets, dev);
- for (i = 0; i < ARRAY_SIZE(op->open_buckets); i++)
- if (op->open_buckets[i]) {
- bch2_open_bucket_put(c,
- c->open_buckets +
- op->open_buckets[i]);
- op->open_buckets[i] = 0;
- }
+ bch2_open_buckets_put(c, &op->open_buckets);
+ return;
+err:
+ keys->top = keys->keys;
+ op->error = ret;
+ goto out;
+}
- if (!(op->flags & BCH_WRITE_DONE))
- continue_at(cl, __bch2_write, op->io_wq);
+static void bch2_write_index(struct closure *cl)
+{
+ struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
+ struct bch_fs *c = op->c;
+
+ __bch2_write_index(op);
- if (!op->error && (op->flags & BCH_WRITE_FLUSH)) {
+ if (!(op->flags & BCH_WRITE_DONE)) {
+ continue_at(cl, __bch2_write, index_update_wq(op));
+ } else if (!op->error && (op->flags & BCH_WRITE_FLUSH)) {
bch2_journal_flush_seq_async(&c->journal,
*op_journal_seq(op),
cl);
}
}
-/**
- * bch_write_discard - discard range of keys
- *
- * Used to implement discard, and to handle when writethrough write hits
- * a write error on the cache device.
- */
-static void bch2_write_discard(struct closure *cl)
+static void bch2_write_endio(struct bio *bio)
{
- struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
- struct bio *bio = &op->bio->bio;
- struct bpos end = op->pos;
+ struct closure *cl = bio->bi_private;
+ struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
+ struct bch_write_bio *wbio = to_wbio(bio);
+ struct bch_write_bio *parent = wbio->split ? wbio->parent : NULL;
+ struct bch_fs *c = wbio->c;
+ struct bch_dev *ca = bch_dev_bkey_exists(c, wbio->dev);
+
+ if (bch2_dev_inum_io_err_on(bio->bi_status, ca,
+ op->pos.inode,
+ op->pos.offset - bio_sectors(bio), /* XXX definitely wrong */
+ "data write error: %s",
+ bch2_blk_status_to_str(bio->bi_status)))
+ set_bit(wbio->dev, op->failed.d);
+
+ if (wbio->have_ioref) {
+ bch2_latency_acct(ca, wbio->submit_time, WRITE);
+ percpu_ref_put(&ca->io_ref);
+ }
- end.offset += bio_sectors(bio);
+ if (wbio->bounce)
+ bch2_bio_free_pages_pool(c, bio);
- op->error = bch2_discard(op->c, op->pos, end, op->version,
- &op->res, NULL, NULL);
+ if (wbio->put_bio)
+ bio_put(bio);
+
+ if (parent)
+ bio_endio(&parent->bio);
+ else if (!(op->flags & BCH_WRITE_SKIP_CLOSURE_PUT))
+ closure_put(cl);
+ else
+ continue_at_nobarrier(cl, bch2_write_index, index_update_wq(op));
}
-/*
- * Convert extents to be inserted to discards after an error:
- */
-static void bch2_write_io_error(struct closure *cl)
+static void init_append_extent(struct bch_write_op *op,
+ struct write_point *wp,
+ struct bversion version,
+ struct bch_extent_crc_unpacked crc)
{
- struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
+ struct bch_fs *c = op->c;
+ struct bkey_i_extent *e;
+ struct open_bucket *ob;
+ unsigned i;
- if (op->flags & BCH_WRITE_DISCARD_ON_ERROR) {
- struct bkey_i *src = bch2_keylist_front(&op->insert_keys);
- struct bkey_i *dst = bch2_keylist_front(&op->insert_keys);
+ BUG_ON(crc.compressed_size > wp->sectors_free);
+ wp->sectors_free -= crc.compressed_size;
+ op->pos.offset += crc.uncompressed_size;
- /*
- * Our data write just errored, which means we've got a bunch
- * of keys to insert that point to data that wasn't
- * successfully written.
- *
- * We don't have to insert those keys but we still have to
- * invalidate that region of the cache - so, if we just strip
- * off all the pointers from the keys we'll accomplish just
- * that.
- */
+ e = bkey_extent_init(op->insert_keys.top);
+ e->k.p = op->pos;
+ e->k.size = crc.uncompressed_size;
+ e->k.version = version;
- while (src != op->insert_keys.top) {
- struct bkey_i *n = bkey_next(src);
+ if (crc.csum_type ||
+ crc.compression_type ||
+ crc.nonce)
+ bch2_extent_crc_append(&e->k_i, crc);
- set_bkey_val_u64s(&src->k, 0);
- src->k.type = KEY_TYPE_DISCARD;
- bkey_copy(dst, src);
+ open_bucket_for_each(c, &wp->ptrs, ob, i) {
+ struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
+ union bch_extent_entry *end =
+ bkey_val_end(bkey_i_to_s(&e->k_i));
- dst = bkey_next(dst);
- src = n;
- }
+ end->ptr = ob->ptr;
+ end->ptr.type = 1 << BCH_EXTENT_ENTRY_ptr;
+ end->ptr.cached = !ca->mi.durability ||
+ (op->flags & BCH_WRITE_CACHED) != 0;
+ end->ptr.offset += ca->mi.bucket_size - ob->sectors_free;
- op->insert_keys.top = dst;
- op->flags |= BCH_WRITE_DISCARD;
- } else {
- /* TODO: We could try to recover from this. */
- while (!bch2_keylist_empty(&op->insert_keys))
- bch2_keylist_pop_front(&op->insert_keys);
+ e->k.u64s++;
- op->error = -EIO;
- op->flags |= BCH_WRITE_DONE;
+ BUG_ON(crc.compressed_size > ob->sectors_free);
+ ob->sectors_free -= crc.compressed_size;
}
- bch2_write_index(cl);
+ bch2_keylist_push(&op->insert_keys);
}
-static void bch2_write_endio(struct bio *bio)
+static struct bio *bch2_write_bio_alloc(struct bch_fs *c,
+ struct write_point *wp,
+ struct bio *src,
+ bool *page_alloc_failed,
+ void *buf)
{
- struct closure *cl = bio->bi_private;
- struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
- struct bch_write_bio *wbio = to_wbio(bio);
- struct bch_fs *c = wbio->c;
- struct bio *orig = wbio->orig;
- struct bch_dev *ca = wbio->ca;
+ struct bch_write_bio *wbio;
+ struct bio *bio;
+ unsigned output_available =
+ min(wp->sectors_free << 9, src->bi_iter.bi_size);
+ unsigned pages = DIV_ROUND_UP(output_available +
+ (buf
+ ? ((unsigned long) buf & (PAGE_SIZE - 1))
+ : 0), PAGE_SIZE);
+
+ pages = min(pages, BIO_MAX_VECS);
+
+ bio = bio_alloc_bioset(GFP_NOIO, pages, &c->bio_write);
+ wbio = wbio_init(bio);
+ wbio->put_bio = true;
+ /* copy WRITE_SYNC flag */
+ wbio->bio.bi_opf = src->bi_opf;
+
+ if (buf) {
+ bch2_bio_map(bio, buf, output_available);
+ return bio;
+ }
- if (bch2_dev_nonfatal_io_err_on(bio->bi_error, ca,
- "data write"))
- set_closure_fn(cl, bch2_write_io_error, index_update_wq(op));
+ wbio->bounce = true;
- if (wbio->have_io_ref)
- percpu_ref_put(&ca->io_ref);
+ /*
+ * We can't use mempool for more than c->sb.encoded_extent_max
+ * worth of pages, but we'd like to allocate more if we can:
+ */
+ bch2_bio_alloc_pages_pool(c, bio,
+ min_t(unsigned, output_available,
+ c->sb.encoded_extent_max << 9));
+
+ if (bio->bi_iter.bi_size < output_available)
+ *page_alloc_failed =
+ bch2_bio_alloc_pages(bio,
+ output_available -
+ bio->bi_iter.bi_size,
+ GFP_NOFS) != 0;
+
+ return bio;
+}
- if (bio->bi_error && orig)
- orig->bi_error = bio->bi_error;
+static int bch2_write_rechecksum(struct bch_fs *c,
+ struct bch_write_op *op,
+ unsigned new_csum_type)
+{
+ struct bio *bio = &op->wbio.bio;
+ struct bch_extent_crc_unpacked new_crc;
+ int ret;
- if (wbio->bounce)
- bch2_bio_free_pages_pool(c, bio);
+ /* bch2_rechecksum_bio() can't encrypt or decrypt data: */
- if (wbio->put_bio)
- bio_put(bio);
+ if (bch2_csum_type_is_encryption(op->crc.csum_type) !=
+ bch2_csum_type_is_encryption(new_csum_type))
+ new_csum_type = op->crc.csum_type;
- if (orig)
- bio_endio(orig);
- else
- closure_put(cl);
-}
+ ret = bch2_rechecksum_bio(c, bio, op->version, op->crc,
+ NULL, &new_crc,
+ op->crc.offset, op->crc.live_size,
+ new_csum_type);
+ if (ret)
+ return ret;
-static struct nonce extent_nonce(struct bversion version,
- unsigned nonce,
- unsigned uncompressed_size,
- unsigned compression_type)
-{
- return (struct nonce) {{
- [0] = cpu_to_le32((nonce << 12) |
- (uncompressed_size << 22)),
- [1] = cpu_to_le32(version.lo),
- [2] = cpu_to_le32(version.lo >> 32),
- [3] = cpu_to_le32(version.hi|
- (compression_type << 24))^BCH_NONCE_EXTENT,
- }};
+ bio_advance(bio, op->crc.offset << 9);
+ bio->bi_iter.bi_size = op->crc.live_size << 9;
+ op->crc = new_crc;
+ return 0;
}
-static void init_append_extent(struct bch_write_op *op,
- unsigned compressed_size,
- unsigned uncompressed_size,
- unsigned compression_type,
- unsigned nonce,
- struct bch_csum csum, unsigned csum_type,
- struct open_bucket *ob)
+static int bch2_write_decrypt(struct bch_write_op *op)
{
- struct bkey_i_extent *e = bkey_extent_init(op->insert_keys.top);
-
- op->pos.offset += uncompressed_size;
- e->k.p = op->pos;
- e->k.size = uncompressed_size;
- e->k.version = op->version;
- bkey_extent_set_cached(&e->k, op->flags & BCH_WRITE_CACHED);
-
- bch2_extent_crc_append(e, compressed_size,
- uncompressed_size,
- compression_type,
- nonce, csum, csum_type);
+ struct bch_fs *c = op->c;
+ struct nonce nonce = extent_nonce(op->version, op->crc);
+ struct bch_csum csum;
- bch2_alloc_sectors_append_ptrs(op->c, e, op->nr_replicas,
- ob, compressed_size);
+ if (!bch2_csum_type_is_encryption(op->crc.csum_type))
+ return 0;
- bkey_extent_set_cached(&e->k, (op->flags & BCH_WRITE_CACHED));
- bch2_keylist_push(&op->insert_keys);
+ /*
+ * If we need to decrypt data in the write path, we'll no longer be able
+ * to verify the existing checksum (poly1305 mac, in this case) after
+ * it's decrypted - this is the last point we'll be able to reverify the
+ * checksum:
+ */
+ csum = bch2_checksum_bio(c, op->crc.csum_type, nonce, &op->wbio.bio);
+ if (bch2_crc_cmp(op->crc.csum, csum))
+ return -EIO;
+
+ bch2_encrypt_bio(c, op->crc.csum_type, nonce, &op->wbio.bio);
+ op->crc.csum_type = 0;
+ op->crc.csum = (struct bch_csum) { 0, 0 };
+ return 0;
}
-static int bch2_write_extent(struct bch_write_op *op,
- struct open_bucket *ob,
- struct bio *orig)
+static enum prep_encoded_ret {
+ PREP_ENCODED_OK,
+ PREP_ENCODED_ERR,
+ PREP_ENCODED_CHECKSUM_ERR,
+ PREP_ENCODED_DO_WRITE,
+} bch2_write_prep_encoded_data(struct bch_write_op *op, struct write_point *wp)
{
struct bch_fs *c = op->c;
- struct bio *bio;
- struct bch_write_bio *wbio;
- unsigned key_to_write_offset = op->insert_keys.top_p -
- op->insert_keys.keys_p;
- struct bkey_i *key_to_write;
- unsigned csum_type = op->csum_type;
- unsigned compression_type = op->compression_type;
- int ret;
+ struct bio *bio = &op->wbio.bio;
- /* don't refetch csum type/compression type */
- barrier();
+ if (!(op->flags & BCH_WRITE_DATA_ENCODED))
+ return PREP_ENCODED_OK;
- /* Need to decompress data? */
- if ((op->flags & BCH_WRITE_DATA_COMPRESSED) &&
- (crc_uncompressed_size(NULL, &op->crc) != op->size ||
- crc_compressed_size(NULL, &op->crc) > ob->sectors_free)) {
- int ret;
+ BUG_ON(bio_sectors(bio) != op->crc.compressed_size);
- ret = bch2_bio_uncompress_inplace(c, orig, op->size, op->crc);
- if (ret)
- return ret;
+ /* Can we just write the entire extent as is? */
+ if (op->crc.uncompressed_size == op->crc.live_size &&
+ op->crc.compressed_size <= wp->sectors_free &&
+ (op->crc.compression_type == op->compression_type ||
+ op->incompressible)) {
+ if (!crc_is_compressed(op->crc) &&
+ op->csum_type != op->crc.csum_type &&
+ bch2_write_rechecksum(c, op, op->csum_type))
+ return PREP_ENCODED_CHECKSUM_ERR;
- op->flags &= ~BCH_WRITE_DATA_COMPRESSED;
+ return PREP_ENCODED_DO_WRITE;
}
- if (op->flags & BCH_WRITE_DATA_COMPRESSED) {
- init_append_extent(op,
- crc_compressed_size(NULL, &op->crc),
- crc_uncompressed_size(NULL, &op->crc),
- op->crc.compression_type,
- op->crc.nonce,
- op->crc.csum,
- op->crc.csum_type,
- ob);
-
- bio = orig;
- wbio = to_wbio(bio);
- wbio->orig = NULL;
- wbio->bounce = false;
- wbio->put_bio = false;
- ret = 0;
- } else if (csum_type != BCH_CSUM_NONE ||
- compression_type != BCH_COMPRESSION_NONE) {
- /* all units here in bytes */
- unsigned total_output = 0, output_available =
- min(ob->sectors_free << 9, orig->bi_iter.bi_size);
- unsigned crc_nonce = bch2_csum_type_is_encryption(csum_type)
- ? op->nonce : 0;
+ /*
+ * If the data is compressed and we couldn't write the entire extent as
+ * is, we have to decompress it:
+ */
+ if (crc_is_compressed(op->crc)) {
struct bch_csum csum;
- struct nonce nonce;
- bio = bio_alloc_bioset(GFP_NOIO,
- DIV_ROUND_UP(output_available, PAGE_SIZE),
- &c->bio_write);
- /*
- * XXX: can't use mempool for more than
- * BCH_COMPRESSED_EXTENT_MAX worth of pages
- */
- bch2_bio_alloc_pages_pool(c, bio, output_available);
+ if (bch2_write_decrypt(op))
+ return PREP_ENCODED_CHECKSUM_ERR;
- /* copy WRITE_SYNC flag */
- bio->bi_opf = orig->bi_opf;
- wbio = to_wbio(bio);
- wbio->orig = NULL;
- wbio->bounce = true;
- wbio->put_bio = true;
-
- do {
- unsigned fragment_compression_type = compression_type;
- size_t dst_len, src_len;
-
- bch2_bio_compress(c, bio, &dst_len,
- orig, &src_len,
- &fragment_compression_type);
-
- BUG_ON(!dst_len || dst_len > bio->bi_iter.bi_size);
- BUG_ON(!src_len || src_len > orig->bi_iter.bi_size);
- BUG_ON(dst_len & (block_bytes(c) - 1));
- BUG_ON(src_len & (block_bytes(c) - 1));
-
- swap(bio->bi_iter.bi_size, dst_len);
- nonce = extent_nonce(op->version,
- crc_nonce,
- src_len >> 9,
- fragment_compression_type),
-
- bch2_encrypt_bio(c, csum_type, nonce, bio);
-
- csum = bch2_checksum_bio(c, csum_type, nonce, bio);
- swap(bio->bi_iter.bi_size, dst_len);
-
- init_append_extent(op,
- dst_len >> 9, src_len >> 9,
- fragment_compression_type,
- crc_nonce, csum, csum_type, ob);
-
- total_output += dst_len;
- bio_advance(bio, dst_len);
- bio_advance(orig, src_len);
- } while (bio->bi_iter.bi_size &&
- orig->bi_iter.bi_size &&
- !bch2_keylist_realloc(&op->insert_keys,
- op->inline_keys,
- ARRAY_SIZE(op->inline_keys),
- BKEY_EXTENT_U64s_MAX));
-
- BUG_ON(total_output > output_available);
-
- memset(&bio->bi_iter, 0, sizeof(bio->bi_iter));
- bio->bi_iter.bi_size = total_output;
+ /* Last point we can still verify checksum: */
+ csum = bch2_checksum_bio(c, op->crc.csum_type,
+ extent_nonce(op->version, op->crc),
+ bio);
+ if (bch2_crc_cmp(op->crc.csum, csum))
+ return PREP_ENCODED_CHECKSUM_ERR;
- /*
- * Free unneeded pages after compressing:
- */
- while (bio->bi_vcnt * PAGE_SIZE >
- round_up(bio->bi_iter.bi_size, PAGE_SIZE))
- mempool_free(bio->bi_io_vec[--bio->bi_vcnt].bv_page,
- &c->bio_bounce_pages);
+ if (bch2_bio_uncompress_inplace(c, bio, &op->crc))
+ return PREP_ENCODED_ERR;
+ }
- ret = orig->bi_iter.bi_size != 0;
- } else {
- bio = bio_next_split(orig, ob->sectors_free, GFP_NOIO,
- &c->bio_write);
+ /*
+ * No longer have compressed data after this point - data might be
+ * encrypted:
+ */
- wbio = to_wbio(bio);
- wbio->orig = NULL;
- wbio->bounce = false;
- wbio->put_bio = bio != orig;
+ /*
+ * If the data is checksummed and we're only writing a subset,
+ * rechecksum and adjust bio to point to currently live data:
+ */
+ if ((op->crc.live_size != op->crc.uncompressed_size ||
+ op->crc.csum_type != op->csum_type) &&
+ bch2_write_rechecksum(c, op, op->csum_type))
+ return PREP_ENCODED_CHECKSUM_ERR;
- init_append_extent(op, bio_sectors(bio), bio_sectors(bio),
- compression_type, 0,
- (struct bch_csum) { 0 }, csum_type, ob);
+ /*
+ * If we want to compress the data, it has to be decrypted:
+ */
+ if ((op->compression_type ||
+ bch2_csum_type_is_encryption(op->crc.csum_type) !=
+ bch2_csum_type_is_encryption(op->csum_type)) &&
+ bch2_write_decrypt(op))
+ return PREP_ENCODED_CHECKSUM_ERR;
- ret = bio != orig;
+ return PREP_ENCODED_OK;
+}
+
+static int bch2_write_extent(struct bch_write_op *op, struct write_point *wp,
+ struct bio **_dst)
+{
+ struct bch_fs *c = op->c;
+ struct bio *src = &op->wbio.bio, *dst = src;
+ struct bvec_iter saved_iter;
+ void *ec_buf;
+ unsigned total_output = 0, total_input = 0;
+ bool bounce = false;
+ bool page_alloc_failed = false;
+ int ret, more = 0;
+
+ BUG_ON(!bio_sectors(src));
+
+ ec_buf = bch2_writepoint_ec_buf(c, wp);
+
+ switch (bch2_write_prep_encoded_data(op, wp)) {
+ case PREP_ENCODED_OK:
+ break;
+ case PREP_ENCODED_ERR:
+ ret = -EIO;
+ goto err;
+ case PREP_ENCODED_CHECKSUM_ERR:
+ goto csum_err;
+ case PREP_ENCODED_DO_WRITE:
+ /* XXX look for bug here */
+ if (ec_buf) {
+ dst = bch2_write_bio_alloc(c, wp, src,
+ &page_alloc_failed,
+ ec_buf);
+ bio_copy_data(dst, src);
+ bounce = true;
+ }
+ init_append_extent(op, wp, op->version, op->crc);
+ goto do_write;
}
- bio->bi_end_io = bch2_write_endio;
- bio->bi_private = &op->cl;
- bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
+ if (ec_buf ||
+ op->compression_type ||
+ (op->csum_type &&
+ !(op->flags & BCH_WRITE_PAGES_STABLE)) ||
+ (bch2_csum_type_is_encryption(op->csum_type) &&
+ !(op->flags & BCH_WRITE_PAGES_OWNED))) {
+ dst = bch2_write_bio_alloc(c, wp, src,
+ &page_alloc_failed,
+ ec_buf);
+ bounce = true;
+ }
- closure_get(bio->bi_private);
+ saved_iter = dst->bi_iter;
- /* might have done a realloc... */
+ do {
+ struct bch_extent_crc_unpacked crc =
+ (struct bch_extent_crc_unpacked) { 0 };
+ struct bversion version = op->version;
+ size_t dst_len, src_len;
+
+ if (page_alloc_failed &&
+ bio_sectors(dst) < wp->sectors_free &&
+ bio_sectors(dst) < c->sb.encoded_extent_max)
+ break;
+
+ BUG_ON(op->compression_type &&
+ (op->flags & BCH_WRITE_DATA_ENCODED) &&
+ bch2_csum_type_is_encryption(op->crc.csum_type));
+ BUG_ON(op->compression_type && !bounce);
+
+ crc.compression_type = op->incompressible
+ ? BCH_COMPRESSION_TYPE_incompressible
+ : op->compression_type
+ ? bch2_bio_compress(c, dst, &dst_len, src, &src_len,
+ op->compression_type)
+ : 0;
+ if (!crc_is_compressed(crc)) {
+ dst_len = min(dst->bi_iter.bi_size, src->bi_iter.bi_size);
+ dst_len = min_t(unsigned, dst_len, wp->sectors_free << 9);
+
+ if (op->csum_type)
+ dst_len = min_t(unsigned, dst_len,
+ c->sb.encoded_extent_max << 9);
+
+ if (bounce) {
+ swap(dst->bi_iter.bi_size, dst_len);
+ bio_copy_data(dst, src);
+ swap(dst->bi_iter.bi_size, dst_len);
+ }
+
+ src_len = dst_len;
+ }
+
+ BUG_ON(!src_len || !dst_len);
- key_to_write = (void *) (op->insert_keys.keys_p + key_to_write_offset);
+ if (bch2_csum_type_is_encryption(op->csum_type)) {
+ if (bversion_zero(version)) {
+ version.lo = atomic64_inc_return(&c->key_version);
+ } else {
+ crc.nonce = op->nonce;
+ op->nonce += src_len >> 9;
+ }
+ }
+
+ if ((op->flags & BCH_WRITE_DATA_ENCODED) &&
+ !crc_is_compressed(crc) &&
+ bch2_csum_type_is_encryption(op->crc.csum_type) ==
+ bch2_csum_type_is_encryption(op->csum_type)) {
+ /*
+ * Note: when we're using rechecksum(), we need to be
+ * checksumming @src because it has all the data our
+ * existing checksum covers - if we bounced (because we
+ * were trying to compress), @dst will only have the
+ * part of the data the new checksum will cover.
+ *
+ * But normally we want to be checksumming post bounce,
+ * because part of the reason for bouncing is so the
+ * data can't be modified (by userspace) while it's in
+ * flight.
+ */
+ if (bch2_rechecksum_bio(c, src, version, op->crc,
+ &crc, &op->crc,
+ src_len >> 9,
+ bio_sectors(src) - (src_len >> 9),
+ op->csum_type))
+ goto csum_err;
+ } else {
+ if ((op->flags & BCH_WRITE_DATA_ENCODED) &&
+ bch2_rechecksum_bio(c, src, version, op->crc,
+ NULL, &op->crc,
+ src_len >> 9,
+ bio_sectors(src) - (src_len >> 9),
+ op->crc.csum_type))
+ goto csum_err;
+
+ crc.compressed_size = dst_len >> 9;
+ crc.uncompressed_size = src_len >> 9;
+ crc.live_size = src_len >> 9;
+
+ swap(dst->bi_iter.bi_size, dst_len);
+ bch2_encrypt_bio(c, op->csum_type,
+ extent_nonce(version, crc), dst);
+ crc.csum = bch2_checksum_bio(c, op->csum_type,
+ extent_nonce(version, crc), dst);
+ crc.csum_type = op->csum_type;
+ swap(dst->bi_iter.bi_size, dst_len);
+ }
- bch2_check_mark_super(c, key_to_write, false);
+ init_append_extent(op, wp, version, crc);
+
+ if (dst != src)
+ bio_advance(dst, dst_len);
+ bio_advance(src, src_len);
+ total_output += dst_len;
+ total_input += src_len;
+ } while (dst->bi_iter.bi_size &&
+ src->bi_iter.bi_size &&
+ wp->sectors_free &&
+ !bch2_keylist_realloc(&op->insert_keys,
+ op->inline_keys,
+ ARRAY_SIZE(op->inline_keys),
+ BKEY_EXTENT_U64s_MAX));
+
+ more = src->bi_iter.bi_size != 0;
+
+ dst->bi_iter = saved_iter;
+
+ if (dst == src && more) {
+ BUG_ON(total_output != total_input);
+
+ dst = bio_split(src, total_input >> 9,
+ GFP_NOIO, &c->bio_write);
+ wbio_init(dst)->put_bio = true;
+ /* copy WRITE_SYNC flag */
+ dst->bi_opf = src->bi_opf;
+ }
+
+ dst->bi_iter.bi_size = total_output;
+do_write:
+ *_dst = dst;
+ return more;
+csum_err:
+ bch_err(c, "error verifying existing checksum while "
+ "rewriting existing data (memory corruption?)");
+ ret = -EIO;
+err:
+ if (to_wbio(dst)->bounce)
+ bch2_bio_free_pages_pool(c, dst);
+ if (to_wbio(dst)->put_bio)
+ bio_put(dst);
- bch2_submit_wbio_replicas(to_wbio(bio), c, key_to_write);
return ret;
}
{
struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
struct bch_fs *c = op->c;
- struct bio *bio = &op->bio->bio;
- unsigned open_bucket_nr = 0;
- struct open_bucket *b;
+ struct write_point *wp;
+ struct bio *bio;
+ bool skip_put = true;
+ unsigned nofs_flags;
int ret;
- memset(op->open_buckets, 0, sizeof(op->open_buckets));
-
- if (op->flags & BCH_WRITE_DISCARD) {
- op->flags |= BCH_WRITE_DONE;
- bch2_write_discard(cl);
- bio_put(bio);
- continue_at(cl, bch2_write_done, index_update_wq(op));
- }
-
- /*
- * Journal writes are marked REQ_PREFLUSH; if the original write was a
- * flush, it'll wait on the journal write.
- */
- bio->bi_opf &= ~(REQ_PREFLUSH|REQ_FUA);
+ nofs_flags = memalloc_nofs_save();
+again:
+ memset(&op->failed, 0, sizeof(op->failed));
do {
- EBUG_ON(bio->bi_iter.bi_sector != op->pos.offset);
- EBUG_ON(!bio_sectors(bio));
+ struct bkey_i *key_to_write;
+ unsigned key_to_write_offset = op->insert_keys.top_p -
+ op->insert_keys.keys_p;
- if (open_bucket_nr == ARRAY_SIZE(op->open_buckets))
- continue_at(cl, bch2_write_index, index_update_wq(op));
+ /* +1 for possible cache device: */
+ if (op->open_buckets.nr + op->nr_replicas + 1 >
+ ARRAY_SIZE(op->open_buckets.v))
+ goto flush_io;
- /* for the device pointers and 1 for the chksum */
if (bch2_keylist_realloc(&op->insert_keys,
op->inline_keys,
ARRAY_SIZE(op->inline_keys),
BKEY_EXTENT_U64s_MAX))
- continue_at(cl, bch2_write_index, index_update_wq(op));
+ goto flush_io;
+
+ if ((op->flags & BCH_WRITE_FROM_INTERNAL) &&
+ percpu_ref_is_dying(&c->writes)) {
+ ret = -EROFS;
+ goto err;
+ }
- b = bch2_alloc_sectors_start(c, op->wp,
+ /*
+ * The copygc thread is now global, which means it's no longer
+ * freeing up space on specific disks, which means that
+ * allocations for specific disks may hang arbitrarily long:
+ */
+ wp = bch2_alloc_sectors_start(c,
+ op->target,
+ op->opts.erasure_code,
+ op->write_point,
+ &op->devs_have,
op->nr_replicas,
- c->opts.data_replicas_required,
+ op->nr_replicas_required,
op->alloc_reserve,
- (op->flags & BCH_WRITE_ALLOC_NOWAIT) ? NULL : cl);
- EBUG_ON(!b);
-
- if (unlikely(IS_ERR(b))) {
- if (unlikely(PTR_ERR(b) != -EAGAIN)) {
- ret = PTR_ERR(b);
+ op->flags,
+ (op->flags & (BCH_WRITE_ALLOC_NOWAIT|
+ BCH_WRITE_ONLY_SPECIFIED_DEVS)) ? NULL : cl);
+ EBUG_ON(!wp);
+
+ if (unlikely(IS_ERR(wp))) {
+ if (unlikely(PTR_ERR(wp) != -EAGAIN)) {
+ ret = PTR_ERR(wp);
goto err;
}
- /*
- * If we already have some keys, must insert them first
- * before allocating another open bucket. We only hit
- * this case if open_bucket_nr > 1.
- */
- if (!bch2_keylist_empty(&op->insert_keys))
- continue_at(cl, bch2_write_index,
- index_update_wq(op));
+ goto flush_io;
+ }
- /*
- * If we've looped, we're running out of a workqueue -
- * not the bch2_write() caller's context - and we don't
- * want to block the workqueue:
- */
- if (op->flags & BCH_WRITE_LOOPED)
- continue_at(cl, __bch2_write, op->io_wq);
+ /*
+ * It's possible for the allocator to fail, put us on the
+ * freelist waitlist, and then succeed in one of various retry
+ * paths: if that happens, we need to disable the skip_put
+ * optimization because otherwise there won't necessarily be a
+ * barrier before we free the bch_write_op:
+ */
+ if (atomic_read(&cl->remaining) & CLOSURE_WAITING)
+ skip_put = false;
+ bch2_open_bucket_get(c, wp, &op->open_buckets);
+ ret = bch2_write_extent(op, wp, &bio);
+ bch2_alloc_sectors_done(c, wp);
+
+ if (ret < 0)
+ goto err;
+
+ if (ret) {
+ skip_put = false;
+ } else {
/*
- * Otherwise, we do want to block the caller on alloc
- * failure instead of letting it queue up more and more
- * writes:
- * XXX: this technically needs a try_to_freeze() -
- * except that that's not safe because caller may have
- * issued other IO... hmm..
+ * for the skip_put optimization this has to be set
+ * before we submit the bio:
*/
- closure_sync(cl);
- continue;
+ op->flags |= BCH_WRITE_DONE;
}
- BUG_ON(b - c->open_buckets == 0 ||
- b - c->open_buckets > U8_MAX);
- op->open_buckets[open_bucket_nr++] = b - c->open_buckets;
+ bio->bi_end_io = bch2_write_endio;
+ bio->bi_private = &op->cl;
+ bio->bi_opf |= REQ_OP_WRITE;
- ret = bch2_write_extent(op, b, bio);
+ if (!skip_put)
+ closure_get(bio->bi_private);
+ else
+ op->flags |= BCH_WRITE_SKIP_CLOSURE_PUT;
- bch2_alloc_sectors_done(c, op->wp, b);
+ key_to_write = (void *) (op->insert_keys.keys_p +
+ key_to_write_offset);
- if (ret < 0)
- goto err;
+ bch2_submit_wbio_replicas(to_wbio(bio), c, BCH_DATA_user,
+ key_to_write);
} while (ret);
- op->flags |= BCH_WRITE_DONE;
- continue_at(cl, bch2_write_index, index_update_wq(op));
+ if (!skip_put)
+ continue_at(cl, bch2_write_index, index_update_wq(op));
+out:
+ memalloc_nofs_restore(nofs_flags);
+ return;
err:
- if (op->flags & BCH_WRITE_DISCARD_ON_ERROR) {
- /*
- * If we were writing cached data, not doing the write is fine
- * so long as we discard whatever would have been overwritten -
- * then it's equivalent to doing the write and immediately
- * reclaiming it.
- */
-
- bch2_write_discard(cl);
- } else {
- /*
- * Right now we can only error here if we went RO - the
- * allocation failed, but we already checked for -ENOSPC when we
- * got our reservation.
- *
- * XXX capacity might have changed, but we don't check for that
- * yet:
- */
- op->error = ret;
- }
-
+ op->error = ret;
op->flags |= BCH_WRITE_DONE;
+ continue_at(cl, bch2_write_index, index_update_wq(op));
+ goto out;
+flush_io:
/*
- * No reason not to insert keys for whatever data was successfully
- * written (especially for a cmpxchg operation that's moving data
- * around)
+ * If the write can't all be submitted at once, we generally want to
+ * block synchronously as that signals backpressure to the caller.
+ *
+ * However, if we're running out of a workqueue, we can't block here
+ * because we'll be blocking other work items from completing:
*/
- continue_at(cl, !bch2_keylist_empty(&op->insert_keys)
- ? bch2_write_index
- : bch2_write_done, index_update_wq(op));
-}
+ if (current->flags & PF_WQ_WORKER) {
+ continue_at(cl, bch2_write_index, index_update_wq(op));
+ goto out;
+ }
-void bch2_wake_delayed_writes(unsigned long data)
-{
- struct bch_fs *c = (void *) data;
- struct bch_write_op *op;
- unsigned long flags;
+ closure_sync(cl);
- spin_lock_irqsave(&c->foreground_write_pd_lock, flags);
+ if (!bch2_keylist_empty(&op->insert_keys)) {
+ __bch2_write_index(op);
- while ((op = c->write_wait_head)) {
- if (time_after(op->expires, jiffies)) {
- mod_timer(&c->foreground_write_wakeup, op->expires);
- break;
+ if (op->error) {
+ op->flags |= BCH_WRITE_DONE;
+ continue_at_nobarrier(cl, bch2_write_done, NULL);
+ goto out;
}
+ }
- c->write_wait_head = op->next;
- if (!c->write_wait_head)
- c->write_wait_tail = NULL;
+ goto again;
+}
- closure_put(&op->cl);
+static void bch2_write_data_inline(struct bch_write_op *op, unsigned data_len)
+{
+ struct closure *cl = &op->cl;
+ struct bio *bio = &op->wbio.bio;
+ struct bvec_iter iter;
+ struct bkey_i_inline_data *id;
+ unsigned sectors;
+ int ret;
+
+ bch2_check_set_feature(op->c, BCH_FEATURE_inline_data);
+
+ ret = bch2_keylist_realloc(&op->insert_keys, op->inline_keys,
+ ARRAY_SIZE(op->inline_keys),
+ BKEY_U64s + DIV_ROUND_UP(data_len, 8));
+ if (ret) {
+ op->error = ret;
+ goto err;
}
- spin_unlock_irqrestore(&c->foreground_write_pd_lock, flags);
+ sectors = bio_sectors(bio);
+ op->pos.offset += sectors;
+
+ id = bkey_inline_data_init(op->insert_keys.top);
+ id->k.p = op->pos;
+ id->k.version = op->version;
+ id->k.size = sectors;
+
+ iter = bio->bi_iter;
+ iter.bi_size = data_len;
+ memcpy_from_bio(id->v.data, bio, iter);
+
+ while (data_len & 7)
+ id->v.data[data_len++] = '\0';
+ 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;
+
+ continue_at_nobarrier(cl, bch2_write_index, NULL);
+ return;
+err:
+ bch2_write_done(&op->cl);
}
/**
* after the data is written it calls bch_journal, and after the keys have been
* added to the next journal write they're inserted into the btree.
*
- * It inserts the data in op->bio; bi_sector is used for the key offset, and
- * op->inode is used for the key inode.
- *
* If op->discard is true, instead of inserting the data it invalidates the
* region of the cache represented by op->bio and op->inode.
*/
void bch2_write(struct closure *cl)
{
struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
- struct bio *bio = &op->bio->bio;
+ struct bio *bio = &op->wbio.bio;
struct bch_fs *c = op->c;
- u64 inode = op->pos.inode;
+ unsigned data_len;
+
+ BUG_ON(!op->nr_replicas);
+ BUG_ON(!op->write_point.v);
+ BUG_ON(!bkey_cmp(op->pos, POS_MAX));
+
+ op->start_time = local_clock();
+ bch2_keylist_init(&op->insert_keys, op->inline_keys);
+ wbio_init(bio)->put_bio = false;
+
+ if (bio_sectors(bio) & (c->opts.block_size - 1)) {
+ bch_err_inum_ratelimited(c, op->pos.inode,
+ "misaligned write");
+ op->error = -EIO;
+ goto err;
+ }
if (c->opts.nochanges ||
!percpu_ref_tryget(&c->writes)) {
- __bcache_io_error(c, "read only");
op->error = -EROFS;
- bch2_disk_reservation_put(c, &op->res);
+ goto err;
+ }
+
+ bch2_increment_clock(c, bio_sectors(bio), WRITE);
+
+ data_len = min_t(u64, bio->bi_iter.bi_size,
+ op->new_i_size - (op->pos.offset << 9));
+
+ if (c->opts.inline_data &&
+ data_len <= min(block_bytes(c) / 2, 1024U)) {
+ bch2_write_data_inline(op, data_len);
+ return;
+ }
+
+ continue_at_nobarrier(cl, __bch2_write, NULL);
+ return;
+err:
+ bch2_disk_reservation_put(c, &op->res);
+
+ if (op->end_io) {
+ EBUG_ON(cl->parent);
+ closure_debug_destroy(cl);
+ op->end_io(op);
+ } else {
closure_return(cl);
}
+}
- if (bversion_zero(op->version) &&
- bch2_csum_type_is_encryption(op->csum_type))
- op->version.lo =
- atomic64_inc_return(&c->key_version) + 1;
+/* Cache promotion on read */
+
+struct promote_op {
+ struct closure cl;
+ struct rcu_head rcu;
+ u64 start_time;
+
+ struct rhash_head hash;
+ struct bpos pos;
+
+ struct migrate_write write;
+ struct bio_vec bi_inline_vecs[0]; /* must be last */
+};
+
+static const struct rhashtable_params bch_promote_params = {
+ .head_offset = offsetof(struct promote_op, hash),
+ .key_offset = offsetof(struct promote_op, pos),
+ .key_len = sizeof(struct bpos),
+};
+
+static inline bool should_promote(struct bch_fs *c, struct bkey_s_c k,
+ struct bpos pos,
+ struct bch_io_opts opts,
+ unsigned flags)
+{
+ if (!(flags & BCH_READ_MAY_PROMOTE))
+ return false;
+
+ if (!opts.promote_target)
+ return false;
+
+ if (bch2_bkey_has_target(c, k, opts.promote_target))
+ return false;
+
+ if (bch2_target_congested(c, opts.promote_target)) {
+ /* XXX trace this */
+ return false;
+ }
+
+ if (rhashtable_lookup_fast(&c->promote_table, &pos,
+ bch_promote_params))
+ return false;
+
+ return true;
+}
+
+static void promote_free(struct bch_fs *c, struct promote_op *op)
+{
+ int ret;
+
+ ret = rhashtable_remove_fast(&c->promote_table, &op->hash,
+ bch_promote_params);
+ BUG_ON(ret);
+ percpu_ref_put(&c->writes);
+ kfree_rcu(op, rcu);
+}
+
+static void promote_done(struct closure *cl)
+{
+ struct promote_op *op =
+ container_of(cl, struct promote_op, cl);
+ struct bch_fs *c = op->write.op.c;
+
+ bch2_time_stats_update(&c->times[BCH_TIME_data_promote],
+ op->start_time);
+
+ bch2_bio_free_pages_pool(c, &op->write.op.wbio.bio);
+ promote_free(c, op);
+}
+
+static void promote_start(struct promote_op *op, struct bch_read_bio *rbio)
+{
+ struct bch_fs *c = rbio->c;
+ struct closure *cl = &op->cl;
+ struct bio *bio = &op->write.op.wbio.bio;
+
+ trace_promote(&rbio->bio);
+
+ /* we now own pages: */
+ BUG_ON(!rbio->bounce);
+ BUG_ON(rbio->bio.bi_vcnt > bio->bi_max_vecs);
+
+ memcpy(bio->bi_io_vec, rbio->bio.bi_io_vec,
+ sizeof(struct bio_vec) * rbio->bio.bi_vcnt);
+ swap(bio->bi_vcnt, rbio->bio.bi_vcnt);
+
+ bch2_migrate_read_done(&op->write, rbio);
+
+ closure_init(cl, NULL);
+ closure_call(&op->write.op.cl, bch2_write, c->btree_update_wq, cl);
+ closure_return_with_destructor(cl, promote_done);
+}
+
+static struct promote_op *__promote_alloc(struct bch_fs *c,
+ enum btree_id btree_id,
+ struct bkey_s_c k,
+ struct bpos pos,
+ struct extent_ptr_decoded *pick,
+ struct bch_io_opts opts,
+ unsigned sectors,
+ struct bch_read_bio **rbio)
+{
+ struct promote_op *op = NULL;
+ struct bio *bio;
+ unsigned pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
+ int ret;
+
+ if (!percpu_ref_tryget(&c->writes))
+ return NULL;
+
+ op = kzalloc(sizeof(*op) + sizeof(struct bio_vec) * pages, GFP_NOIO);
+ if (!op)
+ goto err;
+
+ op->start_time = local_clock();
+ op->pos = pos;
+
+ /*
+ * We don't use the mempool here because extents that aren't
+ * checksummed or compressed can be too big for the mempool:
+ */
+ *rbio = kzalloc(sizeof(struct bch_read_bio) +
+ sizeof(struct bio_vec) * pages,
+ GFP_NOIO);
+ if (!*rbio)
+ goto err;
+
+ rbio_init(&(*rbio)->bio, opts);
+ bio_init(&(*rbio)->bio, (*rbio)->bio.bi_inline_vecs, pages);
+
+ if (bch2_bio_alloc_pages(&(*rbio)->bio, sectors << 9,
+ GFP_NOIO))
+ goto err;
+
+ (*rbio)->bounce = true;
+ (*rbio)->split = true;
+ (*rbio)->kmalloc = true;
+
+ if (rhashtable_lookup_insert_fast(&c->promote_table, &op->hash,
+ bch_promote_params))
+ goto err;
+
+ bio = &op->write.op.wbio.bio;
+ bio_init(bio, bio->bi_inline_vecs, pages);
+
+ ret = bch2_migrate_write_init(c, &op->write,
+ writepoint_hashed((unsigned long) current),
+ opts,
+ DATA_PROMOTE,
+ (struct data_opts) {
+ .target = opts.promote_target,
+ .nr_replicas = 1,
+ },
+ btree_id, k);
+ BUG_ON(ret);
+
+ return op;
+err:
+ if (*rbio)
+ bio_free_pages(&(*rbio)->bio);
+ kfree(*rbio);
+ *rbio = NULL;
+ kfree(op);
+ percpu_ref_put(&c->writes);
+ return NULL;
+}
- if (!(op->flags & BCH_WRITE_DISCARD))
- bch2_increment_clock(c, bio_sectors(bio), WRITE);
+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)
+{
+ 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
+ ? max(pick->crc.compressed_size, pick->crc.live_size)
+ : bvec_iter_sectors(iter);
+ struct bpos pos = promote_full
+ ? bkey_start_pos(k.k)
+ : POS(k.k->p.inode, iter.bi_sector);
+ struct promote_op *promote;
+
+ if (!should_promote(c, k, pos, opts, flags))
+ return NULL;
+
+ promote = __promote_alloc(c,
+ k.k->type == KEY_TYPE_reflink_v
+ ? BTREE_ID_reflink
+ : BTREE_ID_extents,
+ k, pos, pick, opts, sectors, rbio);
+ if (!promote)
+ return NULL;
+
+ *bounce = true;
+ *read_full = promote_full;
+ return promote;
+}
- /* Don't call bch2_next_delay() if rate is >= 1 GB/sec */
+/* Read */
- if (c->foreground_write_ratelimit_enabled &&
- c->foreground_write_pd.rate.rate < (1 << 30) &&
- !(op->flags & BCH_WRITE_DISCARD) && op->wp->throttle) {
- unsigned long flags;
- u64 delay;
+#define READ_RETRY_AVOID 1
+#define READ_RETRY 2
+#define READ_ERR 3
- spin_lock_irqsave(&c->foreground_write_pd_lock, flags);
- bch2_ratelimit_increment(&c->foreground_write_pd.rate,
- bio->bi_iter.bi_size);
+enum rbio_context {
+ RBIO_CONTEXT_NULL,
+ RBIO_CONTEXT_HIGHPRI,
+ RBIO_CONTEXT_UNBOUND,
+};
- delay = bch2_ratelimit_delay(&c->foreground_write_pd.rate);
+static inline struct bch_read_bio *
+bch2_rbio_parent(struct bch_read_bio *rbio)
+{
+ return rbio->split ? rbio->parent : rbio;
+}
- if (delay >= HZ / 100) {
- trace_write_throttle(c, inode, bio, delay);
+__always_inline
+static void bch2_rbio_punt(struct bch_read_bio *rbio, work_func_t fn,
+ enum rbio_context context,
+ struct workqueue_struct *wq)
+{
+ if (context <= rbio->context) {
+ fn(&rbio->work);
+ } else {
+ rbio->work.func = fn;
+ rbio->context = context;
+ queue_work(wq, &rbio->work);
+ }
+}
- closure_get(&op->cl); /* list takes a ref */
+static inline struct bch_read_bio *bch2_rbio_free(struct bch_read_bio *rbio)
+{
+ BUG_ON(rbio->bounce && !rbio->split);
- op->expires = jiffies + delay;
- op->next = NULL;
+ if (rbio->promote)
+ promote_free(rbio->c, rbio->promote);
+ rbio->promote = NULL;
- if (c->write_wait_tail)
- c->write_wait_tail->next = op;
- else
- c->write_wait_head = op;
- c->write_wait_tail = op;
+ if (rbio->bounce)
+ bch2_bio_free_pages_pool(rbio->c, &rbio->bio);
- if (!timer_pending(&c->foreground_write_wakeup))
- mod_timer(&c->foreground_write_wakeup,
- op->expires);
+ if (rbio->split) {
+ struct bch_read_bio *parent = rbio->parent;
- spin_unlock_irqrestore(&c->foreground_write_pd_lock,
- flags);
- continue_at(cl, __bch2_write, index_update_wq(op));
- }
+ if (rbio->kmalloc)
+ kfree(rbio);
+ else
+ bio_put(&rbio->bio);
- spin_unlock_irqrestore(&c->foreground_write_pd_lock, flags);
+ rbio = parent;
}
- continue_at_nobarrier(cl, __bch2_write, NULL);
+ return rbio;
}
-void bch2_write_op_init(struct bch_write_op *op, struct bch_fs *c,
- struct bch_write_bio *bio, struct disk_reservation res,
- struct write_point *wp, struct bpos pos,
- u64 *journal_seq, unsigned flags)
+/*
+ * Only called on a top level bch_read_bio to complete an entire read request,
+ * not a split:
+ */
+static void bch2_rbio_done(struct bch_read_bio *rbio)
{
- EBUG_ON(res.sectors && !res.nr_replicas);
-
- op->c = c;
- op->io_wq = index_update_wq(op);
- op->bio = bio;
- op->written = 0;
- op->error = 0;
- op->flags = flags;
- op->csum_type = bch2_data_checksum_type(c);
- op->compression_type = c->opts.compression;
- op->nr_replicas = res.nr_replicas;
- op->alloc_reserve = RESERVE_NONE;
- op->nonce = 0;
- op->pos = pos;
- op->version = ZERO_VERSION;
- op->res = res;
- op->wp = wp;
-
- if (journal_seq) {
- op->journal_seq_p = journal_seq;
- op->flags |= BCH_WRITE_JOURNAL_SEQ_PTR;
- } else {
- op->journal_seq = 0;
- }
+ if (rbio->start_time)
+ bch2_time_stats_update(&rbio->c->times[BCH_TIME_data_read],
+ rbio->start_time);
+ bio_endio(&rbio->bio);
+}
- op->index_update_fn = bch2_write_index_default;
+static void bch2_read_retry_nodecode(struct bch_fs *c, struct bch_read_bio *rbio,
+ struct bvec_iter bvec_iter,
+ struct bch_io_failures *failed,
+ unsigned flags)
+{
+ struct btree_trans trans;
+ struct btree_iter iter;
+ struct bkey_buf sk;
+ struct bkey_s_c k;
+ int ret;
- bch2_keylist_init(&op->insert_keys,
- op->inline_keys,
- ARRAY_SIZE(op->inline_keys));
+ flags &= ~BCH_READ_LAST_FRAGMENT;
+ flags |= BCH_READ_MUST_CLONE;
- if (version_stress_test(c))
- get_random_bytes(&op->version, sizeof(op->version));
-}
+ bch2_bkey_buf_init(&sk);
+ bch2_trans_init(&trans, c, 0, 0);
-/* Discard */
+ bch2_trans_iter_init(&trans, &iter, rbio->data_btree,
+ rbio->read_pos, BTREE_ITER_SLOTS);
+retry:
+ rbio->bio.bi_status = 0;
-/* bch_discard - discard a range of keys from start_key to end_key.
- * @c filesystem
- * @start_key pointer to start location
- * NOTE: discard starts at bkey_start_offset(start_key)
- * @end_key pointer to end location
- * NOTE: discard ends at KEY_OFFSET(end_key)
- * @version version of discard (0ULL if none)
- *
- * Returns:
- * 0 on success
- * <0 on error
- *
- * XXX: this needs to be refactored with inode_truncate, or more
- * appropriately inode_truncate should call this
- */
-int bch2_discard(struct bch_fs *c, struct bpos start,
- struct bpos end, struct bversion version,
- struct disk_reservation *disk_res,
- struct extent_insert_hook *hook,
- u64 *journal_seq)
-{
- return bch2_btree_delete_range(c, BTREE_ID_EXTENTS, start, end, version,
- disk_res, hook, journal_seq);
-}
+ k = bch2_btree_iter_peek_slot(&iter);
+ if (bkey_err(k))
+ goto err;
-/* Cache promotion on read */
+ bch2_bkey_buf_reassemble(&sk, c, k);
+ k = bkey_i_to_s_c(sk.k);
+ bch2_trans_unlock(&trans);
-struct cache_promote_op {
- struct closure cl;
- struct migrate_write write;
- struct bio_vec bi_inline_vecs[0]; /* must be last */
-};
+ if (!bch2_bkey_matches_ptr(c, k,
+ rbio->pick.ptr,
+ rbio->data_pos.offset -
+ rbio->pick.crc.offset)) {
+ /* extent we wanted to read no longer exists: */
+ rbio->hole = true;
+ goto out;
+ }
-/* Read */
+ ret = __bch2_read_extent(&trans, rbio, bvec_iter,
+ rbio->read_pos,
+ rbio->data_btree,
+ k, 0, failed, flags);
+ if (ret == READ_RETRY)
+ goto retry;
+ if (ret)
+ goto err;
+out:
+ bch2_rbio_done(rbio);
+ bch2_trans_iter_exit(&trans, &iter);
+ bch2_trans_exit(&trans);
+ bch2_bkey_buf_exit(&sk, c);
+ return;
+err:
+ rbio->bio.bi_status = BLK_STS_IOERR;
+ goto out;
+}
-static int bio_checksum_uncompress(struct bch_fs *c,
- struct bch_read_bio *rbio)
+static void bch2_rbio_retry(struct work_struct *work)
{
- struct bio *src = &rbio->bio;
- struct bio *dst = &bch2_rbio_parent(rbio)->bio;
- struct bvec_iter dst_iter = rbio->parent_iter;
- struct nonce nonce = extent_nonce(rbio->version,
- rbio->crc.nonce,
- crc_uncompressed_size(NULL, &rbio->crc),
- rbio->crc.compression_type);
- struct bch_csum csum;
- int ret = 0;
-
- /*
- * reset iterator for checksumming and copying bounced data: here we've
- * set rbio->compressed_size to the amount of data we actually read,
- * which was not necessarily the full extent if we were only bouncing
- * in order to promote
- */
- if (rbio->bounce) {
- src->bi_iter.bi_size = crc_compressed_size(NULL, &rbio->crc) << 9;
- src->bi_iter.bi_idx = 0;
- src->bi_iter.bi_bvec_done = 0;
- } else {
- src->bi_iter = rbio->parent_iter;
- }
+ struct bch_read_bio *rbio =
+ container_of(work, struct bch_read_bio, work);
+ struct bch_fs *c = rbio->c;
+ struct bvec_iter iter = rbio->bvec_iter;
+ unsigned flags = rbio->flags;
+ subvol_inum inum = {
+ .subvol = rbio->subvol,
+ .inum = rbio->read_pos.inode,
+ };
+ struct bch_io_failures failed = { .nr = 0 };
- csum = bch2_checksum_bio(c, rbio->crc.csum_type, nonce, src);
- if (bch2_dev_nonfatal_io_err_on(bch2_crc_cmp(rbio->crc.csum, csum),
- rbio->ca,
- "data checksum error, inode %llu offset %llu: expected %0llx%0llx got %0llx%0llx (type %u)",
- rbio->inode, (u64) rbio->parent_iter.bi_sector << 9,
- rbio->crc.csum.hi, rbio->crc.csum.lo, csum.hi, csum.lo,
- rbio->crc.csum_type))
- ret = -EIO;
+ trace_read_retry(&rbio->bio);
- /*
- * If there was a checksum error, still copy the data back - unless it
- * was compressed, we don't want to decompress bad data:
- */
- if (rbio->crc.compression_type != BCH_COMPRESSION_NONE) {
- if (!ret) {
- bch2_encrypt_bio(c, rbio->crc.csum_type, nonce, src);
- ret = bch2_bio_uncompress(c, src, dst,
- dst_iter, rbio->crc);
- if (ret)
- __bcache_io_error(c, "decompression error");
- }
- } else if (rbio->bounce) {
- bio_advance(src, rbio->crc.offset << 9);
+ if (rbio->retry == READ_RETRY_AVOID)
+ bch2_mark_io_failure(&failed, &rbio->pick);
- /* don't need to decrypt the entire bio: */
- BUG_ON(src->bi_iter.bi_size < dst_iter.bi_size);
- src->bi_iter.bi_size = dst_iter.bi_size;
+ rbio->bio.bi_status = 0;
- nonce = nonce_add(nonce, rbio->crc.offset << 9);
+ rbio = bch2_rbio_free(rbio);
- bch2_encrypt_bio(c, rbio->crc.csum_type,
- nonce, src);
+ flags |= BCH_READ_IN_RETRY;
+ flags &= ~BCH_READ_MAY_PROMOTE;
- bio_copy_data_iter(dst, dst_iter,
- src, src->bi_iter);
+ if (flags & BCH_READ_NODECODE) {
+ bch2_read_retry_nodecode(c, rbio, iter, &failed, flags);
} else {
- bch2_encrypt_bio(c, rbio->crc.csum_type, nonce, src);
- }
+ flags &= ~BCH_READ_LAST_FRAGMENT;
+ flags |= BCH_READ_MUST_CLONE;
- return ret;
+ __bch2_read(c, rbio, iter, inum, &failed, flags);
+ }
}
-static void bch2_rbio_free(struct bch_read_bio *rbio)
+static void bch2_rbio_error(struct bch_read_bio *rbio, int retry,
+ blk_status_t error)
{
- struct bch_fs *c = rbio->c;
- struct bio *bio = &rbio->bio;
+ rbio->retry = retry;
- BUG_ON(rbio->ca);
- BUG_ON(!rbio->split);
+ if (rbio->flags & BCH_READ_IN_RETRY)
+ return;
- if (rbio->promote)
- kfree(rbio->promote);
- if (rbio->bounce)
- bch2_bio_free_pages_pool(c, bio);
+ if (retry == READ_ERR) {
+ rbio = bch2_rbio_free(rbio);
- bio_put(bio);
+ rbio->bio.bi_status = error;
+ bch2_rbio_done(rbio);
+ } else {
+ bch2_rbio_punt(rbio, bch2_rbio_retry,
+ RBIO_CONTEXT_UNBOUND, system_unbound_wq);
+ }
}
-static void bch2_rbio_done(struct bch_read_bio *rbio)
+static int __bch2_rbio_narrow_crcs(struct btree_trans *trans,
+ struct bch_read_bio *rbio)
{
- struct bio *orig = &bch2_rbio_parent(rbio)->bio;
-
- percpu_ref_put(&rbio->ca->io_ref);
- rbio->ca = NULL;
-
- if (rbio->split) {
- if (rbio->bio.bi_error)
- orig->bi_error = rbio->bio.bi_error;
-
- bio_endio(orig);
- bch2_rbio_free(rbio);
- } else {
- if (rbio->promote)
- kfree(rbio->promote);
+ struct bch_fs *c = rbio->c;
+ u64 data_offset = rbio->data_pos.offset - rbio->pick.crc.offset;
+ struct bch_extent_crc_unpacked new_crc;
+ struct btree_iter iter;
+ struct bkey_i *new;
+ struct bkey_s_c k;
+ int ret = 0;
- orig->bi_end_io = rbio->orig_bi_end_io;
- bio_endio_nodec(orig);
+ if (crc_is_compressed(rbio->pick.crc))
+ return 0;
+
+ bch2_trans_iter_init(trans, &iter, rbio->data_btree, rbio->data_pos,
+ BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
+ k = bch2_btree_iter_peek_slot(&iter);
+ if ((ret = bkey_err(k)))
+ goto out;
+
+ if (bversion_cmp(k.k->version, rbio->version) ||
+ !bch2_bkey_matches_ptr(c, k, rbio->pick.ptr, data_offset))
+ goto out;
+
+ /* Extent was merged? */
+ if (bkey_start_offset(k.k) < data_offset ||
+ k.k->p.offset > data_offset + rbio->pick.crc.uncompressed_size)
+ goto out;
+
+ if (bch2_rechecksum_bio(c, &rbio->bio, rbio->version,
+ rbio->pick.crc, NULL, &new_crc,
+ bkey_start_offset(k.k) - data_offset, k.k->size,
+ rbio->pick.crc.csum_type)) {
+ bch_err(c, "error verifying existing checksum while narrowing checksum (memory corruption?)");
+ ret = 0;
+ goto out;
}
-}
-static void bch2_rbio_error(struct bch_read_bio *rbio, int error)
-{
- bch2_rbio_parent(rbio)->bio.bi_error = error;
- bch2_rbio_done(rbio);
-}
+ /*
+ * going to be temporarily appending another checksum entry:
+ */
+ new = bch2_trans_kmalloc(trans, bkey_bytes(k.k) +
+ sizeof(struct bch_extent_crc128));
+ if ((ret = PTR_ERR_OR_ZERO(new)))
+ goto out;
-static void bch2_rbio_retry(struct bch_fs *c, struct bch_read_bio *rbio)
-{
- unsigned long flags;
+ bkey_reassemble(new, k);
- percpu_ref_put(&rbio->ca->io_ref);
- rbio->ca = NULL;
+ if (!bch2_bkey_narrow_crcs(new, new_crc))
+ goto out;
- spin_lock_irqsave(&c->read_retry_lock, flags);
- bio_list_add(&c->read_retry_list, &rbio->bio);
- spin_unlock_irqrestore(&c->read_retry_lock, flags);
- queue_work(c->wq, &c->read_retry_work);
+ ret = bch2_trans_update(trans, &iter, new,
+ BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
+out:
+ bch2_trans_iter_exit(trans, &iter);
+ return ret;
}
-static void cache_promote_done(struct closure *cl)
+static noinline void bch2_rbio_narrow_crcs(struct bch_read_bio *rbio)
{
- struct cache_promote_op *op =
- container_of(cl, struct cache_promote_op, cl);
-
- bch2_bio_free_pages_pool(op->write.op.c, &op->write.wbio.bio);
- kfree(op);
+ bch2_trans_do(rbio->c, NULL, NULL, BTREE_INSERT_NOFAIL,
+ __bch2_rbio_narrow_crcs(&trans, rbio));
}
/* Inner part that may run in process context */
{
struct bch_read_bio *rbio =
container_of(work, struct bch_read_bio, work);
- struct bch_fs *c = rbio->c;
- int ret;
+ struct bch_fs *c = rbio->c;
+ struct bch_dev *ca = bch_dev_bkey_exists(c, rbio->pick.ptr.dev);
+ struct bio *src = &rbio->bio;
+ struct bio *dst = &bch2_rbio_parent(rbio)->bio;
+ struct bvec_iter dst_iter = rbio->bvec_iter;
+ struct bch_extent_crc_unpacked crc = rbio->pick.crc;
+ struct nonce nonce = extent_nonce(rbio->version, crc);
+ unsigned nofs_flags;
+ struct bch_csum csum;
- ret = bio_checksum_uncompress(c, rbio);
- if (ret) {
- /*
- * Checksum error: if the bio wasn't bounced, we may have been
- * reading into buffers owned by userspace (that userspace can
- * scribble over) - retry the read, bouncing it this time:
- */
- if (!rbio->bounce && (rbio->flags & BCH_READ_USER_MAPPED)) {
- rbio->flags |= BCH_READ_FORCE_BOUNCE;
- bch2_rbio_retry(c, rbio);
- } else {
- bch2_rbio_error(rbio, -EIO);
- }
- return;
- }
+ nofs_flags = memalloc_nofs_save();
- if (rbio->promote) {
- struct cache_promote_op *promote = rbio->promote;
- struct closure *cl = &promote->cl;
+ /* Reset iterator for checksumming and copying bounced data: */
+ if (rbio->bounce) {
+ src->bi_iter.bi_size = crc.compressed_size << 9;
+ src->bi_iter.bi_idx = 0;
+ src->bi_iter.bi_bvec_done = 0;
+ } else {
+ src->bi_iter = rbio->bvec_iter;
+ }
- BUG_ON(!rbio->split || !rbio->bounce);
+ csum = bch2_checksum_bio(c, crc.csum_type, nonce, src);
+ if (bch2_crc_cmp(csum, rbio->pick.crc.csum))
+ goto csum_err;
- trace_promote(&rbio->bio);
+ /*
+ * XXX
+ * We need to rework the narrow_crcs path to deliver the read completion
+ * first, and then punt to a different workqueue, otherwise we're
+ * holding up reads while doing btree updates which is bad for memory
+ * reclaim.
+ */
+ if (unlikely(rbio->narrow_crcs))
+ bch2_rbio_narrow_crcs(rbio);
- /* we now own pages: */
- swap(promote->write.wbio.bio.bi_vcnt, rbio->bio.bi_vcnt);
- rbio->promote = NULL;
+ if (rbio->flags & BCH_READ_NODECODE)
+ goto nodecode;
- bch2_rbio_done(rbio);
+ /* Adjust crc to point to subset of data we want: */
+ crc.offset += rbio->offset_into_extent;
+ crc.live_size = bvec_iter_sectors(rbio->bvec_iter);
- closure_init(cl, &c->cl);
- closure_call(&promote->write.op.cl, bch2_write, c->wq, cl);
- closure_return_with_destructor(cl, cache_promote_done);
+ if (crc_is_compressed(crc)) {
+ bch2_encrypt_bio(c, crc.csum_type, nonce, src);
+ if (bch2_bio_uncompress(c, src, dst, dst_iter, crc))
+ goto decompression_err;
} else {
+ /* don't need to decrypt the entire bio: */
+ nonce = nonce_add(nonce, crc.offset << 9);
+ bio_advance(src, crc.offset << 9);
+
+ BUG_ON(src->bi_iter.bi_size < dst_iter.bi_size);
+ src->bi_iter.bi_size = dst_iter.bi_size;
+
+ bch2_encrypt_bio(c, crc.csum_type, nonce, src);
+
+ if (rbio->bounce) {
+ struct bvec_iter src_iter = src->bi_iter;
+ bio_copy_data_iter(dst, &dst_iter, src, &src_iter);
+ }
+ }
+
+ if (rbio->promote) {
+ /*
+ * Re encrypt data we decrypted, so it's consistent with
+ * rbio->crc:
+ */
+ bch2_encrypt_bio(c, crc.csum_type, nonce, src);
+ promote_start(rbio->promote, rbio);
+ rbio->promote = NULL;
+ }
+nodecode:
+ if (likely(!(rbio->flags & BCH_READ_IN_RETRY))) {
+ rbio = bch2_rbio_free(rbio);
bch2_rbio_done(rbio);
}
+out:
+ memalloc_nofs_restore(nofs_flags);
+ return;
+csum_err:
+ /*
+ * Checksum error: if the bio wasn't bounced, we may have been
+ * reading into buffers owned by userspace (that userspace can
+ * scribble over) - retry the read, bouncing it this time:
+ */
+ if (!rbio->bounce && (rbio->flags & BCH_READ_USER_MAPPED)) {
+ rbio->flags |= BCH_READ_MUST_BOUNCE;
+ bch2_rbio_error(rbio, READ_RETRY, BLK_STS_IOERR);
+ goto out;
+ }
+
+ bch2_dev_inum_io_error(ca, rbio->read_pos.inode, (u64) rbio->bvec_iter.bi_sector,
+ "data checksum error: expected %0llx:%0llx got %0llx:%0llx (type %u)",
+ rbio->pick.crc.csum.hi, rbio->pick.crc.csum.lo,
+ csum.hi, csum.lo, crc.csum_type);
+ bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
+ goto out;
+decompression_err:
+ bch_err_inum_ratelimited(c, rbio->read_pos.inode,
+ "decompression error");
+ bch2_rbio_error(rbio, READ_ERR, BLK_STS_IOERR);
+ goto out;
}
static void bch2_read_endio(struct bio *bio)
{
struct bch_read_bio *rbio =
container_of(bio, struct bch_read_bio, bio);
- struct bch_fs *c = rbio->c;
+ struct bch_fs *c = rbio->c;
+ struct bch_dev *ca = bch_dev_bkey_exists(c, rbio->pick.ptr.dev);
+ struct workqueue_struct *wq = NULL;
+ enum rbio_context context = RBIO_CONTEXT_NULL;
+
+ if (rbio->have_ioref) {
+ bch2_latency_acct(ca, rbio->submit_time, READ);
+ percpu_ref_put(&ca->io_ref);
+ }
- if (bch2_dev_nonfatal_io_err_on(bio->bi_error, rbio->ca, "data read")) {
- /* XXX: retry IO errors when we have another replica */
- bch2_rbio_error(rbio, bio->bi_error);
+ if (!rbio->split)
+ rbio->bio.bi_end_io = rbio->end_io;
+
+ if (bch2_dev_inum_io_err_on(bio->bi_status, ca,
+ rbio->read_pos.inode,
+ rbio->read_pos.offset,
+ "data read error: %s",
+ bch2_blk_status_to_str(bio->bi_status))) {
+ bch2_rbio_error(rbio, READ_RETRY_AVOID, bio->bi_status);
return;
}
- if (rbio->ptr.cached &&
+ if (rbio->pick.ptr.cached &&
(((rbio->flags & BCH_READ_RETRY_IF_STALE) && race_fault()) ||
- ptr_stale(rbio->ca, &rbio->ptr))) {
+ ptr_stale(ca, &rbio->pick.ptr))) {
atomic_long_inc(&c->read_realloc_races);
if (rbio->flags & BCH_READ_RETRY_IF_STALE)
- bch2_rbio_retry(c, rbio);
+ bch2_rbio_error(rbio, READ_RETRY, BLK_STS_AGAIN);
else
- bch2_rbio_error(rbio, -EINTR);
+ bch2_rbio_error(rbio, READ_ERR, BLK_STS_AGAIN);
return;
}
- if (rbio->crc.compression_type ||
- bch2_csum_type_is_encryption(rbio->crc.csum_type))
- queue_work(system_unbound_wq, &rbio->work);
- else if (rbio->crc.csum_type)
- queue_work(system_highpri_wq, &rbio->work);
- else
- __bch2_read_endio(&rbio->work);
+ if (rbio->narrow_crcs ||
+ crc_is_compressed(rbio->pick.crc) ||
+ bch2_csum_type_is_encryption(rbio->pick.crc.csum_type))
+ context = RBIO_CONTEXT_UNBOUND, wq = system_unbound_wq;
+ else if (rbio->pick.crc.csum_type)
+ context = RBIO_CONTEXT_HIGHPRI, wq = system_highpri_wq;
+
+ bch2_rbio_punt(rbio, __bch2_read_endio, context, wq);
}
-static bool should_promote(struct bch_fs *c,
- struct extent_pick_ptr *pick, unsigned flags)
+int __bch2_read_indirect_extent(struct btree_trans *trans,
+ unsigned *offset_into_extent,
+ struct bkey_buf *orig_k)
{
- if (!(flags & BCH_READ_PROMOTE))
- return false;
+ struct btree_iter iter;
+ struct bkey_s_c k;
+ u64 reflink_offset;
+ int ret;
- if (percpu_ref_is_dying(&c->writes))
- return false;
+ reflink_offset = le64_to_cpu(bkey_i_to_reflink_p(orig_k->k)->v.idx) +
+ *offset_into_extent;
+
+ bch2_trans_iter_init(trans, &iter, BTREE_ID_reflink,
+ POS(0, reflink_offset),
+ BTREE_ITER_SLOTS);
+ k = bch2_btree_iter_peek_slot(&iter);
+ ret = bkey_err(k);
+ if (ret)
+ goto err;
+
+ if (k.k->type != KEY_TYPE_reflink_v &&
+ k.k->type != KEY_TYPE_indirect_inline_data) {
+ bch_err_inum_ratelimited(trans->c, orig_k->k->k.p.inode,
+ "%llu len %u points to nonexistent indirect extent %llu",
+ orig_k->k->k.p.offset,
+ orig_k->k->k.size,
+ reflink_offset);
+ bch2_inconsistent_error(trans->c);
+ ret = -EIO;
+ goto err;
+ }
- return c->fastest_tier &&
- c->fastest_tier < c->tiers + pick->ca->mi.tier;
+ *offset_into_extent = iter.pos.offset - bkey_start_offset(k.k);
+ bch2_bkey_buf_reassemble(orig_k, trans->c, k);
+err:
+ bch2_trans_iter_exit(trans, &iter);
+ return ret;
}
-void bch2_read_extent_iter(struct bch_fs *c, struct bch_read_bio *orig,
- struct bvec_iter iter, struct bkey_s_c k,
- struct extent_pick_ptr *pick, unsigned flags)
+int __bch2_read_extent(struct btree_trans *trans, struct bch_read_bio *orig,
+ struct bvec_iter iter, struct bpos read_pos,
+ enum btree_id data_btree, struct bkey_s_c k,
+ unsigned offset_into_extent,
+ struct bch_io_failures *failed, unsigned flags)
{
- struct bch_read_bio *rbio;
- struct cache_promote_op *promote_op = NULL;
- unsigned skip = iter.bi_sector - bkey_start_offset(k.k);
- bool bounce = false, split, read_full = false;
+ struct bch_fs *c = trans->c;
+ struct extent_ptr_decoded pick;
+ struct bch_read_bio *rbio = NULL;
+ struct bch_dev *ca;
+ struct promote_op *promote = NULL;
+ bool bounce = false, read_full = false, narrow_crcs = false;
+ struct bpos data_pos = bkey_start_pos(k.k);
+ int pick_ret;
+
+ if (bkey_extent_is_inline_data(k.k)) {
+ unsigned bytes = min_t(unsigned, iter.bi_size,
+ bkey_inline_data_bytes(k.k));
+
+ swap(iter.bi_size, bytes);
+ memcpy_to_bio(&orig->bio, iter, bkey_inline_data_p(k));
+ swap(iter.bi_size, bytes);
+ bio_advance_iter(&orig->bio, &iter, bytes);
+ zero_fill_bio_iter(&orig->bio, iter);
+ goto out_read_done;
+ }
- bch2_increment_clock(c, bio_sectors(&orig->bio), READ);
+ pick_ret = bch2_bkey_pick_read_device(c, k, failed, &pick);
- EBUG_ON(bkey_start_offset(k.k) > iter.bi_sector ||
- k.k->p.offset < bvec_iter_end_sector(iter));
+ /* hole or reservation - just zero fill: */
+ if (!pick_ret)
+ goto hole;
- /* only promote if we're not reading from the fastest tier: */
+ if (pick_ret < 0) {
+ bch_err_inum_ratelimited(c, k.k->p.inode,
+ "no device to read from");
+ goto err;
+ }
- /*
- * XXX: multiple promotes can race with each other, wastefully. Keep a
- * list of outstanding promotes?
- */
- if (should_promote(c, pick, flags)) {
+ if (pick_ret > 0)
+ ca = bch_dev_bkey_exists(c, pick.ptr.dev);
+
+ if (flags & BCH_READ_NODECODE) {
/*
- * biovec needs to be big enough to hold decompressed data, if
- * the bch2_write_extent() has to decompress/recompress it:
+ * can happen if we retry, and the extent we were going to read
+ * has been merged in the meantime:
*/
- unsigned sectors =
- max_t(unsigned, k.k->size,
- crc_uncompressed_size(NULL, &pick->crc));
- unsigned pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
-
- promote_op = kmalloc(sizeof(*promote_op) +
- sizeof(struct bio_vec) * pages, GFP_NOIO);
- if (promote_op) {
- struct bio *promote_bio = &promote_op->write.wbio.bio;
-
- bio_init(promote_bio);
- promote_bio->bi_max_vecs = pages;
- promote_bio->bi_io_vec = promote_bio->bi_inline_vecs;
- bounce = true;
- /* could also set read_full */
- }
+ if (pick.crc.compressed_size > orig->bio.bi_vcnt * PAGE_SECTORS)
+ goto hole;
+
+ iter.bi_size = pick.crc.compressed_size << 9;
+ goto get_bio;
}
- /*
- * note: if compression_type and crc_type both == none, then
- * compressed/uncompressed size is zero
- */
- if (pick->crc.compression_type != BCH_COMPRESSION_NONE ||
- (pick->crc.csum_type != BCH_CSUM_NONE &&
- (bvec_iter_sectors(iter) != crc_uncompressed_size(NULL, &pick->crc) ||
- (bch2_csum_type_is_encryption(pick->crc.csum_type) &&
+ if (!(flags & BCH_READ_LAST_FRAGMENT) ||
+ bio_flagged(&orig->bio, BIO_CHAIN))
+ flags |= BCH_READ_MUST_CLONE;
+
+ narrow_crcs = !(flags & BCH_READ_IN_RETRY) &&
+ bch2_can_narrow_extent_crcs(k, pick.crc);
+
+ if (narrow_crcs && (flags & BCH_READ_USER_MAPPED))
+ flags |= BCH_READ_MUST_BOUNCE;
+
+ EBUG_ON(offset_into_extent + bvec_iter_sectors(iter) > k.k->size);
+
+ if (crc_is_compressed(pick.crc) ||
+ (pick.crc.csum_type != BCH_CSUM_none &&
+ (bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
+ (bch2_csum_type_is_encryption(pick.crc.csum_type) &&
(flags & BCH_READ_USER_MAPPED)) ||
- (flags & BCH_READ_FORCE_BOUNCE)))) {
+ (flags & BCH_READ_MUST_BOUNCE)))) {
read_full = true;
bounce = true;
}
- if (bounce) {
- unsigned sectors = read_full
- ? (crc_compressed_size(NULL, &pick->crc) ?: k.k->size)
- : bvec_iter_sectors(iter);
-
- rbio = container_of(bio_alloc_bioset(GFP_NOIO,
- DIV_ROUND_UP(sectors, PAGE_SECTORS),
- &c->bio_read_split),
- struct bch_read_bio, bio);
+ if (orig->opts.promote_target)
+ promote = promote_alloc(c, iter, k, &pick, orig->opts, flags,
+ &rbio, &bounce, &read_full);
+
+ if (!read_full) {
+ EBUG_ON(crc_is_compressed(pick.crc));
+ EBUG_ON(pick.crc.csum_type &&
+ (bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
+ bvec_iter_sectors(iter) != pick.crc.live_size ||
+ pick.crc.offset ||
+ offset_into_extent));
+
+ data_pos.offset += offset_into_extent;
+ pick.ptr.offset += pick.crc.offset +
+ offset_into_extent;
+ offset_into_extent = 0;
+ pick.crc.compressed_size = bvec_iter_sectors(iter);
+ pick.crc.uncompressed_size = bvec_iter_sectors(iter);
+ pick.crc.offset = 0;
+ pick.crc.live_size = bvec_iter_sectors(iter);
+ offset_into_extent = 0;
+ }
+get_bio:
+ if (rbio) {
+ /*
+ * promote already allocated bounce rbio:
+ * promote needs to allocate a bio big enough for uncompressing
+ * data in the write path, but we're not going to use it all
+ * here:
+ */
+ EBUG_ON(rbio->bio.bi_iter.bi_size <
+ pick.crc.compressed_size << 9);
+ rbio->bio.bi_iter.bi_size =
+ pick.crc.compressed_size << 9;
+ } else if (bounce) {
+ unsigned sectors = pick.crc.compressed_size;
+
+ rbio = rbio_init(bio_alloc_bioset(GFP_NOIO,
+ DIV_ROUND_UP(sectors, PAGE_SECTORS),
+ &c->bio_read_split),
+ orig->opts);
bch2_bio_alloc_pages_pool(c, &rbio->bio, sectors << 9);
- split = true;
- } else if (!(flags & BCH_READ_MAY_REUSE_BIO) ||
- !(flags & BCH_READ_IS_LAST)) {
+ rbio->bounce = true;
+ rbio->split = true;
+ } else if (flags & BCH_READ_MUST_CLONE) {
/*
* Have to clone if there were any splits, due to error
* reporting issues (if a split errored, and retrying didn't
* from the whole bio, in which case we don't want to retry and
* lose the error)
*/
- rbio = container_of(bio_clone_fast(&orig->bio,
- GFP_NOIO, &c->bio_read_split),
- struct bch_read_bio, bio);
+ rbio = rbio_init(bio_clone_fast(&orig->bio, GFP_NOIO,
+ &c->bio_read_split),
+ orig->opts);
rbio->bio.bi_iter = iter;
- split = true;
+ rbio->split = true;
} else {
rbio = orig;
rbio->bio.bi_iter = iter;
- split = false;
- BUG_ON(bio_flagged(&rbio->bio, BIO_CHAIN));
+ EBUG_ON(bio_flagged(&rbio->bio, BIO_CHAIN));
}
- if (!(flags & BCH_READ_IS_LAST))
- __bio_inc_remaining(&orig->bio);
+ EBUG_ON(bio_sectors(&rbio->bio) != pick.crc.compressed_size);
- if (split)
+ rbio->c = c;
+ rbio->submit_time = local_clock();
+ if (rbio->split)
rbio->parent = orig;
else
- rbio->orig_bi_end_io = orig->bio.bi_end_io;
- rbio->parent_iter = iter;
-
+ rbio->end_io = orig->bio.bi_end_io;
+ rbio->bvec_iter = iter;
+ rbio->offset_into_extent= offset_into_extent;
rbio->flags = flags;
- rbio->bounce = bounce;
- rbio->split = split;
- rbio->c = c;
- rbio->ca = pick->ca;
- rbio->ptr = pick->ptr;
- rbio->crc = pick->crc;
- /*
- * crc.compressed_size will be 0 if there wasn't any checksum
- * information, also we need to stash the original size of the bio if we
- * bounced (which isn't necessarily the original key size, if we bounced
- * only for promoting)
- */
- rbio->crc._compressed_size = bio_sectors(&rbio->bio) - 1;
+ rbio->have_ioref = pick_ret > 0 && bch2_dev_get_ioref(ca, READ);
+ rbio->narrow_crcs = narrow_crcs;
+ rbio->hole = 0;
+ rbio->retry = 0;
+ rbio->context = 0;
+ /* XXX: only initialize this if needed */
+ rbio->devs_have = bch2_bkey_devs(k);
+ rbio->pick = pick;
+ rbio->subvol = orig->subvol;
+ rbio->read_pos = read_pos;
+ rbio->data_btree = data_btree;
+ rbio->data_pos = data_pos;
rbio->version = k.k->version;
- rbio->promote = promote_op;
- rbio->inode = k.k->p.inode;
- INIT_WORK(&rbio->work, __bch2_read_endio);
+ rbio->promote = promote;
+ INIT_WORK(&rbio->work, NULL);
- rbio->bio.bi_bdev = pick->ca->disk_sb.bdev;
rbio->bio.bi_opf = orig->bio.bi_opf;
- rbio->bio.bi_iter.bi_sector = pick->ptr.offset;
+ rbio->bio.bi_iter.bi_sector = pick.ptr.offset;
rbio->bio.bi_end_io = bch2_read_endio;
- if (promote_op) {
- struct bio *promote_bio = &promote_op->write.wbio.bio;
-
- promote_bio->bi_iter = rbio->bio.bi_iter;
- memcpy(promote_bio->bi_io_vec, rbio->bio.bi_io_vec,
- sizeof(struct bio_vec) * rbio->bio.bi_vcnt);
-
- bch2_migrate_write_init(c, &promote_op->write,
- &c->promote_write_point,
- k, NULL,
- BCH_WRITE_ALLOC_NOWAIT|
- BCH_WRITE_CACHED);
- promote_op->write.promote = true;
-
- if (rbio->crc.compression_type) {
- promote_op->write.op.flags |= BCH_WRITE_DATA_COMPRESSED;
- promote_op->write.op.crc = rbio->crc;
- promote_op->write.op.size = k.k->size;
- } else if (read_full) {
- /*
- * Adjust bio to correspond to _live_ portion of @k -
- * which might be less than what we're actually reading:
- */
- bio_advance(promote_bio, rbio->crc.offset << 9);
- BUG_ON(bio_sectors(promote_bio) < k.k->size);
- promote_bio->bi_iter.bi_size = k.k->size << 9;
- } else {
- /*
- * Set insert pos to correspond to what we're actually
- * reading:
- */
- promote_op->write.op.pos.offset = iter.bi_sector;
+ if (rbio->bounce)
+ trace_read_bounce(&rbio->bio);
+
+ bch2_increment_clock(c, bio_sectors(&rbio->bio), READ);
+
+ /*
+ * If it's being moved internally, we don't want to flag it as a cache
+ * hit:
+ */
+ if (pick.ptr.cached && !(flags & BCH_READ_NODECODE))
+ bch2_bucket_io_time_reset(trans, pick.ptr.dev,
+ PTR_BUCKET_NR(ca, &pick.ptr), READ);
+
+ if (!(flags & (BCH_READ_IN_RETRY|BCH_READ_LAST_FRAGMENT))) {
+ bio_inc_remaining(&orig->bio);
+ trace_read_split(&orig->bio);
+ }
+
+ if (!rbio->pick.idx) {
+ if (!rbio->have_ioref) {
+ bch_err_inum_ratelimited(c, k.k->p.inode,
+ "no device to read from");
+ bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
+ goto out;
+ }
+
+ this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_user],
+ 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);
+ } else {
+ /* Attempting reconstruct read: */
+ if (bch2_ec_read_extent(c, rbio)) {
+ bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
+ goto out;
}
- promote_bio->bi_iter.bi_sector =
- promote_op->write.op.pos.offset;
+ if (likely(!(flags & BCH_READ_IN_RETRY)))
+ bio_endio(&rbio->bio);
}
+out:
+ if (likely(!(flags & BCH_READ_IN_RETRY))) {
+ return 0;
+ } else {
+ int ret;
- /* _after_ promete stuff has looked at rbio->crc.offset */
- if (read_full)
- rbio->crc.offset += skip;
- else
- rbio->bio.bi_iter.bi_sector += skip;
+ rbio->context = RBIO_CONTEXT_UNBOUND;
+ bch2_read_endio(&rbio->bio);
- rbio->submit_time_us = local_clock_us();
+ ret = rbio->retry;
+ rbio = bch2_rbio_free(rbio);
- if (bounce)
- trace_read_bounce(&rbio->bio);
+ if (ret == READ_RETRY_AVOID) {
+ bch2_mark_io_failure(failed, &pick);
+ ret = READ_RETRY;
+ }
+
+ if (!ret)
+ goto out_read_done;
+
+ return ret;
+ }
+
+err:
+ if (flags & BCH_READ_IN_RETRY)
+ return READ_ERR;
- if (!(flags & BCH_READ_IS_LAST))
- trace_read_split(&rbio->bio);
+ orig->bio.bi_status = BLK_STS_IOERR;
+ goto out_read_done;
- generic_make_request(&rbio->bio);
+hole:
+ /*
+ * won't normally happen in the BCH_READ_NODECODE
+ * (bch2_move_extent()) path, but if we retry and the extent we wanted
+ * to read no longer exists we have to signal that:
+ */
+ if (flags & BCH_READ_NODECODE)
+ orig->hole = true;
+
+ zero_fill_bio_iter(&orig->bio, iter);
+out_read_done:
+ if (flags & BCH_READ_LAST_FRAGMENT)
+ bch2_rbio_done(orig);
+ return 0;
}
-static void bch2_read_iter(struct bch_fs *c, struct bch_read_bio *rbio,
- struct bvec_iter bvec_iter, u64 inode,
- unsigned flags)
+void __bch2_read(struct bch_fs *c, struct bch_read_bio *rbio,
+ struct bvec_iter bvec_iter, subvol_inum inum,
+ struct bch_io_failures *failed, unsigned flags)
{
- struct bio *bio = &rbio->bio;
+ struct btree_trans trans;
struct btree_iter iter;
+ struct bkey_buf sk;
struct bkey_s_c k;
+ u32 snapshot;
int ret;
- for_each_btree_key_with_holes(&iter, c, BTREE_ID_EXTENTS,
- POS(inode, bvec_iter.bi_sector), k) {
- BKEY_PADDED(k) tmp;
- struct extent_pick_ptr pick;
- unsigned bytes, sectors;
- bool is_last;
+ BUG_ON(flags & BCH_READ_NODECODE);
+
+ bch2_bkey_buf_init(&sk);
+ bch2_trans_init(&trans, c, 0, 0);
+retry:
+ bch2_trans_begin(&trans);
+ iter = (struct btree_iter) { NULL };
+
+ ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
+ if (ret)
+ goto err;
+
+ bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents,
+ SPOS(inum.inum, bvec_iter.bi_sector, snapshot),
+ BTREE_ITER_SLOTS|BTREE_ITER_FILTER_SNAPSHOTS);
+ while (1) {
+ unsigned bytes, sectors, offset_into_extent;
+ enum btree_id data_btree = BTREE_ID_extents;
/*
- * Unlock the iterator while the btree node's lock is still in
- * cache, before doing the IO:
+ * read_extent -> io_time_reset may cause a transaction restart
+ * without returning an error, we need to check for that here:
*/
- bkey_reassemble(&tmp.k, k);
- k = bkey_i_to_s_c(&tmp.k);
- bch2_btree_iter_unlock(&iter);
-
- bch2_extent_pick_ptr(c, k, &pick);
- if (IS_ERR(pick.ca)) {
- bcache_io_error(c, bio, "no device to read from");
- bio_endio(bio);
- return;
+ if (!bch2_trans_relock(&trans)) {
+ ret = -EINTR;
+ break;
}
- sectors = min_t(u64, k.k->p.offset,
- bvec_iter_end_sector(bvec_iter)) -
- bvec_iter.bi_sector;
- bytes = sectors << 9;
- is_last = bytes == bvec_iter.bi_size;
- swap(bvec_iter.bi_size, bytes);
+ bch2_btree_iter_set_pos(&iter,
+ POS(inum.inum, bvec_iter.bi_sector));
- if (is_last)
- flags |= BCH_READ_IS_LAST;
+ k = bch2_btree_iter_peek_slot(&iter);
+ ret = bkey_err(k);
+ if (ret)
+ break;
- if (pick.ca) {
- PTR_BUCKET(pick.ca, &pick.ptr)->read_prio =
- c->prio_clock[READ].hand;
+ offset_into_extent = iter.pos.offset -
+ bkey_start_offset(k.k);
+ sectors = k.k->size - offset_into_extent;
- bch2_read_extent_iter(c, rbio, bvec_iter,
- k, &pick, flags);
+ bch2_bkey_buf_reassemble(&sk, c, k);
- flags &= ~BCH_READ_MAY_REUSE_BIO;
- } else {
- zero_fill_bio_iter(bio, bvec_iter);
+ ret = bch2_read_indirect_extent(&trans, &data_btree,
+ &offset_into_extent, &sk);
+ if (ret)
+ break;
- if (is_last)
- bio_endio(bio);
- }
+ k = bkey_i_to_s_c(sk.k);
- if (is_last)
- return;
+ /*
+ * With indirect extents, the amount of data to read is the min
+ * of the original extent and the indirect extent:
+ */
+ sectors = min(sectors, k.k->size - offset_into_extent);
+
+ /*
+ * Unlock the iterator while the btree node's lock is still in
+ * cache, before doing the IO:
+ */
+ bch2_trans_unlock(&trans);
+ bytes = min(sectors, bvec_iter_sectors(bvec_iter)) << 9;
swap(bvec_iter.bi_size, bytes);
- bio_advance_iter(bio, &bvec_iter, bytes);
- }
- /*
- * If we get here, it better have been because there was an error
- * reading a btree node
- */
- ret = bch2_btree_iter_unlock(&iter);
- BUG_ON(!ret);
- bcache_io_error(c, bio, "btree IO error %i", ret);
- bio_endio(bio);
-}
+ if (bvec_iter.bi_size == bytes)
+ flags |= BCH_READ_LAST_FRAGMENT;
-void bch2_read(struct bch_fs *c, struct bch_read_bio *bio, u64 inode)
-{
- bch2_read_iter(c, bio, bio->bio.bi_iter, inode,
- BCH_READ_RETRY_IF_STALE|
- BCH_READ_PROMOTE|
- BCH_READ_MAY_REUSE_BIO|
- BCH_READ_USER_MAPPED);
-}
+ ret = __bch2_read_extent(&trans, rbio, bvec_iter, iter.pos,
+ data_btree, k,
+ offset_into_extent, failed, flags);
+ if (ret)
+ break;
-/**
- * bch_read_retry - re-submit a bio originally from bch2_read()
- */
-static void bch2_read_retry(struct bch_fs *c, struct bch_read_bio *rbio)
-{
- struct bch_read_bio *parent = bch2_rbio_parent(rbio);
- struct bvec_iter iter = rbio->parent_iter;
- unsigned flags = rbio->flags;
- u64 inode = rbio->inode;
+ if (flags & BCH_READ_LAST_FRAGMENT)
+ break;
- trace_read_retry(&rbio->bio);
+ swap(bvec_iter.bi_size, bytes);
+ bio_advance_iter(&rbio->bio, &bvec_iter, bytes);
- if (rbio->split)
- bch2_rbio_free(rbio);
- else
- rbio->bio.bi_end_io = rbio->orig_bi_end_io;
+ ret = btree_trans_too_many_iters(&trans);
+ if (ret)
+ break;
+ }
+err:
+ bch2_trans_iter_exit(&trans, &iter);
- bch2_read_iter(c, parent, iter, inode, flags);
-}
+ if (ret == -EINTR || ret == READ_RETRY || ret == READ_RETRY_AVOID)
+ goto retry;
-void bch2_read_retry_work(struct work_struct *work)
-{
- struct bch_fs *c = container_of(work, struct bch_fs,
- read_retry_work);
- struct bch_read_bio *rbio;
- struct bio *bio;
- unsigned long flags;
+ bch2_trans_exit(&trans);
+ bch2_bkey_buf_exit(&sk, c);
- while (1) {
- spin_lock_irqsave(&c->read_retry_lock, flags);
- bio = bio_list_pop(&c->read_retry_list);
- spin_unlock_irqrestore(&c->read_retry_lock, flags);
+ if (ret) {
+ bch_err_inum_ratelimited(c, inum.inum,
+ "read error %i from btree lookup", ret);
+ rbio->bio.bi_status = BLK_STS_IOERR;
+ bch2_rbio_done(rbio);
+ }
+}
- if (!bio)
- break;
+void bch2_fs_io_exit(struct bch_fs *c)
+{
+ if (c->promote_table.tbl)
+ rhashtable_destroy(&c->promote_table);
+ mempool_exit(&c->bio_bounce_pages);
+ bioset_exit(&c->bio_write);
+ bioset_exit(&c->bio_read_split);
+ bioset_exit(&c->bio_read);
+}
- rbio = container_of(bio, struct bch_read_bio, bio);
- bch2_read_retry(c, rbio);
- }
+int bch2_fs_io_init(struct bch_fs *c)
+{
+ 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),
+ BIOSET_NEED_BVECS) ||
+ bioset_init(&c->bio_write, 1, offsetof(struct bch_write_bio, bio),
+ BIOSET_NEED_BVECS) ||
+ mempool_init_page_pool(&c->bio_bounce_pages,
+ max_t(unsigned,
+ c->opts.btree_node_size,
+ c->sb.encoded_extent_max) /
+ PAGE_SECTORS, 0) ||
+ rhashtable_init(&c->promote_table, &bch_promote_params))
+ return -ENOMEM;
+
+ return 0;
}