linux/fs/bcachefs/journal_io.c

1678 lines
43 KiB
C
Raw Normal View History

// SPDX-License-Identifier: GPL-2.0
#include "bcachefs.h"
#include "alloc_foreground.h"
#include "btree_io.h"
#include "btree_update_interior.h"
#include "buckets.h"
#include "checksum.h"
#include "disk_groups.h"
#include "error.h"
#include "io.h"
#include "journal.h"
#include "journal_io.h"
#include "journal_reclaim.h"
bcachefs: Don't require flush/fua on every journal write This patch adds a flag to journal entries which, if set, indicates that they weren't done as flush/fua writes. - non flush/fua journal writes don't update last_seq (i.e. they don't free up space in the journal), thus the journal free space calculations now check whether nonflush journal writes are currently allowed (i.e. are we low on free space, or would doing a flush write free up a lot of space in the journal) - write_delay_ms, the user configurable option for when open journal entries are automatically written, is now interpreted as the max delay between flush journal writes (default 1 second). - bch2_journal_flush_seq_async is changed to ensure a flush write >= the requested sequence number has happened - journal read/replay must now ignore, and blacklist, any journal entries newer than the most recent flush entry in the journal. Also, the way the read_entire_journal option is handled has been improved; struct journal_replay now has an entry, 'ignore', for entries that were read but should not be used. - assorted refactoring and improvements related to journal read in journal_io.c and recovery.c Previously, we'd have to issue a flush/fua write every time we accumulated a full journal entry - typically the bucket size. Now we need to issue them much less frequently: when an fsync is requested, or it's been more than write_delay_ms since the last flush, or when we need to free up space in the journal. This is a significant performance improvement on many write heavy workloads. Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com> Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
2020-11-14 22:59:58 +08:00
#include "journal_seq_blacklist.h"
#include "replicas.h"
#include "trace.h"
bcachefs: Don't require flush/fua on every journal write This patch adds a flag to journal entries which, if set, indicates that they weren't done as flush/fua writes. - non flush/fua journal writes don't update last_seq (i.e. they don't free up space in the journal), thus the journal free space calculations now check whether nonflush journal writes are currently allowed (i.e. are we low on free space, or would doing a flush write free up a lot of space in the journal) - write_delay_ms, the user configurable option for when open journal entries are automatically written, is now interpreted as the max delay between flush journal writes (default 1 second). - bch2_journal_flush_seq_async is changed to ensure a flush write >= the requested sequence number has happened - journal read/replay must now ignore, and blacklist, any journal entries newer than the most recent flush entry in the journal. Also, the way the read_entire_journal option is handled has been improved; struct journal_replay now has an entry, 'ignore', for entries that were read but should not be used. - assorted refactoring and improvements related to journal read in journal_io.c and recovery.c Previously, we'd have to issue a flush/fua write every time we accumulated a full journal entry - typically the bucket size. Now we need to issue them much less frequently: when an fsync is requested, or it's been more than write_delay_ms since the last flush, or when we need to free up space in the journal. This is a significant performance improvement on many write heavy workloads. Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com> Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
2020-11-14 22:59:58 +08:00
static void __journal_replay_free(struct journal_replay *i)
{
list_del(&i->list);
kvpfree(i, offsetof(struct journal_replay, j) +
vstruct_bytes(&i->j));
}
static void journal_replay_free(struct bch_fs *c, struct journal_replay *i)
{
i->ignore = true;
if (!c->opts.read_entire_journal)
__journal_replay_free(i);
}
struct journal_list {
struct closure cl;
struct mutex lock;
struct list_head *head;
int ret;
};
#define JOURNAL_ENTRY_ADD_OK 0
#define JOURNAL_ENTRY_ADD_OUT_OF_RANGE 5
/*
* Given a journal entry we just read, add it to the list of journal entries to
* be replayed:
*/
static int journal_entry_add(struct bch_fs *c, struct bch_dev *ca,
struct bch_extent_ptr entry_ptr,
struct journal_list *jlist, struct jset *j,
bool bad)
{
struct journal_replay *i, *pos, *dup = NULL;
struct bch_extent_ptr *ptr;
struct list_head *where;
size_t bytes = vstruct_bytes(j);
bcachefs: Don't require flush/fua on every journal write This patch adds a flag to journal entries which, if set, indicates that they weren't done as flush/fua writes. - non flush/fua journal writes don't update last_seq (i.e. they don't free up space in the journal), thus the journal free space calculations now check whether nonflush journal writes are currently allowed (i.e. are we low on free space, or would doing a flush write free up a lot of space in the journal) - write_delay_ms, the user configurable option for when open journal entries are automatically written, is now interpreted as the max delay between flush journal writes (default 1 second). - bch2_journal_flush_seq_async is changed to ensure a flush write >= the requested sequence number has happened - journal read/replay must now ignore, and blacklist, any journal entries newer than the most recent flush entry in the journal. Also, the way the read_entire_journal option is handled has been improved; struct journal_replay now has an entry, 'ignore', for entries that were read but should not be used. - assorted refactoring and improvements related to journal read in journal_io.c and recovery.c Previously, we'd have to issue a flush/fua write every time we accumulated a full journal entry - typically the bucket size. Now we need to issue them much less frequently: when an fsync is requested, or it's been more than write_delay_ms since the last flush, or when we need to free up space in the journal. This is a significant performance improvement on many write heavy workloads. Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com> Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
2020-11-14 22:59:58 +08:00
u64 last_seq = 0;
int ret = JOURNAL_ENTRY_ADD_OK;
bcachefs: Don't require flush/fua on every journal write This patch adds a flag to journal entries which, if set, indicates that they weren't done as flush/fua writes. - non flush/fua journal writes don't update last_seq (i.e. they don't free up space in the journal), thus the journal free space calculations now check whether nonflush journal writes are currently allowed (i.e. are we low on free space, or would doing a flush write free up a lot of space in the journal) - write_delay_ms, the user configurable option for when open journal entries are automatically written, is now interpreted as the max delay between flush journal writes (default 1 second). - bch2_journal_flush_seq_async is changed to ensure a flush write >= the requested sequence number has happened - journal read/replay must now ignore, and blacklist, any journal entries newer than the most recent flush entry in the journal. Also, the way the read_entire_journal option is handled has been improved; struct journal_replay now has an entry, 'ignore', for entries that were read but should not be used. - assorted refactoring and improvements related to journal read in journal_io.c and recovery.c Previously, we'd have to issue a flush/fua write every time we accumulated a full journal entry - typically the bucket size. Now we need to issue them much less frequently: when an fsync is requested, or it's been more than write_delay_ms since the last flush, or when we need to free up space in the journal. This is a significant performance improvement on many write heavy workloads. Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com> Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
2020-11-14 22:59:58 +08:00
list_for_each_entry_reverse(i, jlist->head, list) {
if (!JSET_NO_FLUSH(&i->j)) {
last_seq = le64_to_cpu(i->j.last_seq);
break;
}
bcachefs: Don't require flush/fua on every journal write This patch adds a flag to journal entries which, if set, indicates that they weren't done as flush/fua writes. - non flush/fua journal writes don't update last_seq (i.e. they don't free up space in the journal), thus the journal free space calculations now check whether nonflush journal writes are currently allowed (i.e. are we low on free space, or would doing a flush write free up a lot of space in the journal) - write_delay_ms, the user configurable option for when open journal entries are automatically written, is now interpreted as the max delay between flush journal writes (default 1 second). - bch2_journal_flush_seq_async is changed to ensure a flush write >= the requested sequence number has happened - journal read/replay must now ignore, and blacklist, any journal entries newer than the most recent flush entry in the journal. Also, the way the read_entire_journal option is handled has been improved; struct journal_replay now has an entry, 'ignore', for entries that were read but should not be used. - assorted refactoring and improvements related to journal read in journal_io.c and recovery.c Previously, we'd have to issue a flush/fua write every time we accumulated a full journal entry - typically the bucket size. Now we need to issue them much less frequently: when an fsync is requested, or it's been more than write_delay_ms since the last flush, or when we need to free up space in the journal. This is a significant performance improvement on many write heavy workloads. Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com> Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
2020-11-14 22:59:58 +08:00
}
bcachefs: Don't require flush/fua on every journal write This patch adds a flag to journal entries which, if set, indicates that they weren't done as flush/fua writes. - non flush/fua journal writes don't update last_seq (i.e. they don't free up space in the journal), thus the journal free space calculations now check whether nonflush journal writes are currently allowed (i.e. are we low on free space, or would doing a flush write free up a lot of space in the journal) - write_delay_ms, the user configurable option for when open journal entries are automatically written, is now interpreted as the max delay between flush journal writes (default 1 second). - bch2_journal_flush_seq_async is changed to ensure a flush write >= the requested sequence number has happened - journal read/replay must now ignore, and blacklist, any journal entries newer than the most recent flush entry in the journal. Also, the way the read_entire_journal option is handled has been improved; struct journal_replay now has an entry, 'ignore', for entries that were read but should not be used. - assorted refactoring and improvements related to journal read in journal_io.c and recovery.c Previously, we'd have to issue a flush/fua write every time we accumulated a full journal entry - typically the bucket size. Now we need to issue them much less frequently: when an fsync is requested, or it's been more than write_delay_ms since the last flush, or when we need to free up space in the journal. This is a significant performance improvement on many write heavy workloads. Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com> Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
2020-11-14 22:59:58 +08:00
/* Is this entry older than the range we need? */
if (!c->opts.read_entire_journal &&
le64_to_cpu(j->seq) < last_seq) {
ret = JOURNAL_ENTRY_ADD_OUT_OF_RANGE;
goto out;
}
/* Drop entries we don't need anymore */
if (!JSET_NO_FLUSH(j)) {
list_for_each_entry_safe(i, pos, jlist->head, list) {
if (le64_to_cpu(i->j.seq) >= le64_to_cpu(j->last_seq))
break;
bcachefs: Don't require flush/fua on every journal write This patch adds a flag to journal entries which, if set, indicates that they weren't done as flush/fua writes. - non flush/fua journal writes don't update last_seq (i.e. they don't free up space in the journal), thus the journal free space calculations now check whether nonflush journal writes are currently allowed (i.e. are we low on free space, or would doing a flush write free up a lot of space in the journal) - write_delay_ms, the user configurable option for when open journal entries are automatically written, is now interpreted as the max delay between flush journal writes (default 1 second). - bch2_journal_flush_seq_async is changed to ensure a flush write >= the requested sequence number has happened - journal read/replay must now ignore, and blacklist, any journal entries newer than the most recent flush entry in the journal. Also, the way the read_entire_journal option is handled has been improved; struct journal_replay now has an entry, 'ignore', for entries that were read but should not be used. - assorted refactoring and improvements related to journal read in journal_io.c and recovery.c Previously, we'd have to issue a flush/fua write every time we accumulated a full journal entry - typically the bucket size. Now we need to issue them much less frequently: when an fsync is requested, or it's been more than write_delay_ms since the last flush, or when we need to free up space in the journal. This is a significant performance improvement on many write heavy workloads. Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com> Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
2020-11-14 22:59:58 +08:00
journal_replay_free(c, i);
}
}
list_for_each_entry_reverse(i, jlist->head, list) {
if (le64_to_cpu(j->seq) > le64_to_cpu(i->j.seq)) {
where = &i->list;
goto add;
}
}
where = jlist->head;
add:
dup = where->next != jlist->head
? container_of(where->next, struct journal_replay, list)
: NULL;
if (dup && le64_to_cpu(j->seq) != le64_to_cpu(dup->j.seq))
dup = NULL;
/*
* Duplicate journal entries? If so we want the one that didn't have a
* checksum error:
*/
if (dup) {
if (dup->bad) {
/* we'll replace @dup: */
} else if (bad) {
i = dup;
goto found;
} else {
fsck_err_on(bytes != vstruct_bytes(&dup->j) ||
memcmp(j, &dup->j, bytes), c,
"found duplicate but non identical journal entries (seq %llu)",
le64_to_cpu(j->seq));
i = dup;
goto found;
}
}
i = kvpmalloc(offsetof(struct journal_replay, j) + bytes, GFP_KERNEL);
if (!i) {
ret = -ENOMEM;
goto out;
}
i->nr_ptrs = 0;
i->bad = bad;
i->ignore = false;
unsafe_memcpy(&i->j, j, bytes, "embedded variable length struct");
if (dup) {
i->nr_ptrs = dup->nr_ptrs;
memcpy(i->ptrs, dup->ptrs, sizeof(dup->ptrs));
__journal_replay_free(dup);
}
list_add(&i->list, where);
found:
for (ptr = i->ptrs; ptr < i->ptrs + i->nr_ptrs; ptr++) {
if (ptr->dev == ca->dev_idx) {
bch_err(c, "duplicate journal entry %llu on same device",
le64_to_cpu(i->j.seq));
goto out;
}
}
if (i->nr_ptrs >= ARRAY_SIZE(i->ptrs)) {
bch_err(c, "found too many copies of journal entry %llu",
le64_to_cpu(i->j.seq));
goto out;
}
i->ptrs[i->nr_ptrs++] = entry_ptr;
out:
fsck_err:
return ret;
}
static struct nonce journal_nonce(const struct jset *jset)
{
return (struct nonce) {{
[0] = 0,
[1] = ((__le32 *) &jset->seq)[0],
[2] = ((__le32 *) &jset->seq)[1],
[3] = BCH_NONCE_JOURNAL,
}};
}
/* this fills in a range with empty jset_entries: */
static void journal_entry_null_range(void *start, void *end)
{
struct jset_entry *entry;
for (entry = start; entry != end; entry = vstruct_next(entry))
memset(entry, 0, sizeof(*entry));
}
#define JOURNAL_ENTRY_REREAD 5
#define JOURNAL_ENTRY_NONE 6
#define JOURNAL_ENTRY_BAD 7
#define journal_entry_err(c, msg, ...) \
({ \
switch (write) { \
case READ: \
mustfix_fsck_err(c, msg, ##__VA_ARGS__); \
break; \
case WRITE: \
bch_err(c, "corrupt metadata before write:\n" \
msg, ##__VA_ARGS__); \
if (bch2_fs_inconsistent(c)) { \
ret = BCH_FSCK_ERRORS_NOT_FIXED; \
goto fsck_err; \
} \
break; \
} \
true; \
})
#define journal_entry_err_on(cond, c, msg, ...) \
((cond) ? journal_entry_err(c, msg, ##__VA_ARGS__) : false)
#define FSCK_DELETED_KEY 5
static int journal_validate_key(struct bch_fs *c, const char *where,
struct jset_entry *entry,
unsigned level, enum btree_id btree_id,
struct bkey_i *k, const char *type,
unsigned version, int big_endian, int write)
{
void *next = vstruct_next(entry);
const char *invalid;
int ret = 0;
if (journal_entry_err_on(!k->k.u64s, c,
"invalid %s in %s entry offset %zi/%u: k->u64s 0",
type, where,
(u64 *) k - entry->_data,
le16_to_cpu(entry->u64s))) {
entry->u64s = cpu_to_le16((u64 *) k - entry->_data);
journal_entry_null_range(vstruct_next(entry), next);
return FSCK_DELETED_KEY;
}
if (journal_entry_err_on((void *) bkey_next(k) >
(void *) vstruct_next(entry), c,
"invalid %s in %s entry offset %zi/%u: extends past end of journal entry",
type, where,
(u64 *) k - entry->_data,
le16_to_cpu(entry->u64s))) {
entry->u64s = cpu_to_le16((u64 *) k - entry->_data);
journal_entry_null_range(vstruct_next(entry), next);
return FSCK_DELETED_KEY;
}
if (journal_entry_err_on(k->k.format != KEY_FORMAT_CURRENT, c,
"invalid %s in %s entry offset %zi/%u: bad format %u",
type, where,
(u64 *) k - entry->_data,
le16_to_cpu(entry->u64s),
k->k.format)) {
le16_add_cpu(&entry->u64s, -((u16) k->k.u64s));
memmove(k, bkey_next(k), next - (void *) bkey_next(k));
journal_entry_null_range(vstruct_next(entry), next);
return FSCK_DELETED_KEY;
}
if (!write)
bch2_bkey_compat(level, btree_id, version, big_endian,
write, NULL, bkey_to_packed(k));
invalid = bch2_bkey_invalid(c, bkey_i_to_s_c(k),
__btree_node_type(level, btree_id));
if (invalid) {
char buf[160];
bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(k));
mustfix_fsck_err(c, "invalid %s in %s entry offset %zi/%u: %s\n%s",
type, where,
(u64 *) k - entry->_data,
le16_to_cpu(entry->u64s),
invalid, buf);
le16_add_cpu(&entry->u64s, -((u16) k->k.u64s));
memmove(k, bkey_next(k), next - (void *) bkey_next(k));
journal_entry_null_range(vstruct_next(entry), next);
return FSCK_DELETED_KEY;
}
if (write)
bch2_bkey_compat(level, btree_id, version, big_endian,
write, NULL, bkey_to_packed(k));
fsck_err:
return ret;
}
static int journal_entry_btree_keys_validate(struct bch_fs *c,
const char *where,
struct jset_entry *entry,
unsigned version, int big_endian, int write)
{
struct bkey_i *k = entry->start;
while (k != vstruct_last(entry)) {
int ret = journal_validate_key(c, where, entry,
entry->level,
entry->btree_id,
k, "key", version, big_endian, write);
if (ret == FSCK_DELETED_KEY)
continue;
k = bkey_next(k);
}
return 0;
}
static void journal_entry_btree_keys_to_text(struct printbuf *out, struct bch_fs *c,
struct jset_entry *entry)
{
struct bkey_i *k;
pr_buf(out, "btree=%s l=%u ", bch2_btree_ids[entry->btree_id], entry->level);
vstruct_for_each(entry, k)
bch2_bkey_val_to_text(out, c, bkey_i_to_s_c(k));
}
static int journal_entry_btree_root_validate(struct bch_fs *c,
const char *where,
struct jset_entry *entry,
unsigned version, int big_endian, int write)
{
struct bkey_i *k = entry->start;
int ret = 0;
if (journal_entry_err_on(!entry->u64s ||
le16_to_cpu(entry->u64s) != k->k.u64s, c,
"invalid btree root journal entry: wrong number of keys")) {
void *next = vstruct_next(entry);
/*
* we don't want to null out this jset_entry,
* just the contents, so that later we can tell
* we were _supposed_ to have a btree root
*/
entry->u64s = 0;
journal_entry_null_range(vstruct_next(entry), next);
return 0;
}
return journal_validate_key(c, where, entry, 1, entry->btree_id, k,
"btree root", version, big_endian, write);
fsck_err:
return ret;
}
static void journal_entry_btree_root_to_text(struct printbuf *out, struct bch_fs *c,
struct jset_entry *entry)
{
journal_entry_btree_keys_to_text(out, c, entry);
}
static int journal_entry_prio_ptrs_validate(struct bch_fs *c,
const char *where,
struct jset_entry *entry,
unsigned version, int big_endian, int write)
{
/* obsolete, don't care: */
return 0;
}
static void journal_entry_prio_ptrs_to_text(struct printbuf *out, struct bch_fs *c,
struct jset_entry *entry)
{
}
static int journal_entry_blacklist_validate(struct bch_fs *c,
const char *where,
struct jset_entry *entry,
unsigned version, int big_endian, int write)
{
int ret = 0;
if (journal_entry_err_on(le16_to_cpu(entry->u64s) != 1, c,
"invalid journal seq blacklist entry: bad size")) {
journal_entry_null_range(entry, vstruct_next(entry));
}
fsck_err:
return ret;
}
static void journal_entry_blacklist_to_text(struct printbuf *out, struct bch_fs *c,
struct jset_entry *entry)
{
struct jset_entry_blacklist *bl =
container_of(entry, struct jset_entry_blacklist, entry);
pr_buf(out, "seq=%llu", le64_to_cpu(bl->seq));
}
static int journal_entry_blacklist_v2_validate(struct bch_fs *c,
const char *where,
struct jset_entry *entry,
unsigned version, int big_endian, int write)
{
struct jset_entry_blacklist_v2 *bl_entry;
int ret = 0;
if (journal_entry_err_on(le16_to_cpu(entry->u64s) != 2, c,
"invalid journal seq blacklist entry: bad size")) {
journal_entry_null_range(entry, vstruct_next(entry));
goto out;
}
bl_entry = container_of(entry, struct jset_entry_blacklist_v2, entry);
if (journal_entry_err_on(le64_to_cpu(bl_entry->start) >
le64_to_cpu(bl_entry->end), c,
"invalid journal seq blacklist entry: start > end")) {
journal_entry_null_range(entry, vstruct_next(entry));
}
out:
fsck_err:
return ret;
}
static void journal_entry_blacklist_v2_to_text(struct printbuf *out, struct bch_fs *c,
struct jset_entry *entry)
{
struct jset_entry_blacklist_v2 *bl =
container_of(entry, struct jset_entry_blacklist_v2, entry);
pr_buf(out, "start=%llu end=%llu",
le64_to_cpu(bl->start),
le64_to_cpu(bl->end));
}
static int journal_entry_usage_validate(struct bch_fs *c,
const char *where,
struct jset_entry *entry,
unsigned version, int big_endian, int write)
{
struct jset_entry_usage *u =
container_of(entry, struct jset_entry_usage, entry);
unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
int ret = 0;
if (journal_entry_err_on(bytes < sizeof(*u),
c,
"invalid journal entry usage: bad size")) {
journal_entry_null_range(entry, vstruct_next(entry));
return ret;
}
fsck_err:
return ret;
}
static void journal_entry_usage_to_text(struct printbuf *out, struct bch_fs *c,
struct jset_entry *entry)
{
struct jset_entry_usage *u =
container_of(entry, struct jset_entry_usage, entry);
pr_buf(out, "type=%s v=%llu",
bch2_fs_usage_types[u->entry.btree_id],
le64_to_cpu(u->v));
}
static int journal_entry_data_usage_validate(struct bch_fs *c,
const char *where,
struct jset_entry *entry,
unsigned version, int big_endian, int write)
{
struct jset_entry_data_usage *u =
container_of(entry, struct jset_entry_data_usage, entry);
unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
int ret = 0;
if (journal_entry_err_on(bytes < sizeof(*u) ||
bytes < sizeof(*u) + u->r.nr_devs,
c,
"invalid journal entry usage: bad size")) {
journal_entry_null_range(entry, vstruct_next(entry));
return ret;
}
fsck_err:
return ret;
}
static void journal_entry_data_usage_to_text(struct printbuf *out, struct bch_fs *c,
struct jset_entry *entry)
{
struct jset_entry_data_usage *u =
container_of(entry, struct jset_entry_data_usage, entry);
bch2_replicas_entry_to_text(out, &u->r);
pr_buf(out, "=%llu", le64_to_cpu(u->v));
}
static int journal_entry_clock_validate(struct bch_fs *c,
const char *where,
struct jset_entry *entry,
unsigned version, int big_endian, int write)
{
struct jset_entry_clock *clock =
container_of(entry, struct jset_entry_clock, entry);
unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
int ret = 0;
if (journal_entry_err_on(bytes != sizeof(*clock),
c, "invalid journal entry clock: bad size")) {
journal_entry_null_range(entry, vstruct_next(entry));
return ret;
}
if (journal_entry_err_on(clock->rw > 1,
c, "invalid journal entry clock: bad rw")) {
journal_entry_null_range(entry, vstruct_next(entry));
return ret;
}
fsck_err:
return ret;
}
static void journal_entry_clock_to_text(struct printbuf *out, struct bch_fs *c,
struct jset_entry *entry)
{
struct jset_entry_clock *clock =
container_of(entry, struct jset_entry_clock, entry);
pr_buf(out, "%s=%llu", clock->rw ? "write" : "read", le64_to_cpu(clock->time));
}
static int journal_entry_dev_usage_validate(struct bch_fs *c,
const char *where,
struct jset_entry *entry,
unsigned version, int big_endian, int write)
{
struct jset_entry_dev_usage *u =
container_of(entry, struct jset_entry_dev_usage, entry);
unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
unsigned expected = sizeof(*u);
unsigned dev;
int ret = 0;
if (journal_entry_err_on(bytes < expected,
c, "invalid journal entry dev usage: bad size (%u < %u)",
bytes, expected)) {
journal_entry_null_range(entry, vstruct_next(entry));
return ret;
}
dev = le32_to_cpu(u->dev);
if (journal_entry_err_on(!bch2_dev_exists2(c, dev),
c, "invalid journal entry dev usage: bad dev")) {
journal_entry_null_range(entry, vstruct_next(entry));
return ret;
}
if (journal_entry_err_on(u->pad,
c, "invalid journal entry dev usage: bad pad")) {
journal_entry_null_range(entry, vstruct_next(entry));
return ret;
}
fsck_err:
return ret;
}
static void journal_entry_dev_usage_to_text(struct printbuf *out, struct bch_fs *c,
struct jset_entry *entry)
{
struct jset_entry_dev_usage *u =
container_of(entry, struct jset_entry_dev_usage, entry);
unsigned i, nr_types = jset_entry_dev_usage_nr_types(u);
pr_buf(out, "dev=%u", le32_to_cpu(u->dev));
for (i = 0; i < nr_types; i++) {
if (i < BCH_DATA_NR)
pr_buf(out, " %s", bch2_data_types[i]);
else
pr_buf(out, " (unknown data type %u)", i);
pr_buf(out, ": buckets=%llu sectors=%llu fragmented=%llu",
le64_to_cpu(u->d[i].buckets),
le64_to_cpu(u->d[i].sectors),
le64_to_cpu(u->d[i].fragmented));
}
pr_buf(out, " buckets_ec: %llu buckets_unavailable: %llu",
le64_to_cpu(u->buckets_ec),
le64_to_cpu(u->buckets_unavailable));
}
static int journal_entry_log_validate(struct bch_fs *c,
const char *where,
struct jset_entry *entry,
unsigned version, int big_endian, int write)
{
return 0;
}
static void journal_entry_log_to_text(struct printbuf *out, struct bch_fs *c,
struct jset_entry *entry)
{
struct jset_entry_log *l = container_of(entry, struct jset_entry_log, entry);
unsigned bytes = vstruct_bytes(entry) - offsetof(struct jset_entry_log, d);
bch_scnmemcpy(out, l->d, strnlen(l->d, bytes));
}
struct jset_entry_ops {
int (*validate)(struct bch_fs *, const char *,
struct jset_entry *, unsigned, int, int);
void (*to_text)(struct printbuf *, struct bch_fs *, struct jset_entry *);
};
static const struct jset_entry_ops bch2_jset_entry_ops[] = {
#define x(f, nr) \
[BCH_JSET_ENTRY_##f] = (struct jset_entry_ops) { \
.validate = journal_entry_##f##_validate, \
.to_text = journal_entry_##f##_to_text, \
},
BCH_JSET_ENTRY_TYPES()
#undef x
};
int bch2_journal_entry_validate(struct bch_fs *c, const char *where,
struct jset_entry *entry,
unsigned version, int big_endian, int write)
{
return entry->type < BCH_JSET_ENTRY_NR
? bch2_jset_entry_ops[entry->type].validate(c, where, entry,
version, big_endian, write)
: 0;
}
void bch2_journal_entry_to_text(struct printbuf *out, struct bch_fs *c,
struct jset_entry *entry)
{
if (entry->type < BCH_JSET_ENTRY_NR) {
pr_buf(out, "%s: ", bch2_jset_entry_types[entry->type]);
bch2_jset_entry_ops[entry->type].to_text(out, c, entry);
} else {
pr_buf(out, "(unknown type %u)", entry->type);
}
}
static int jset_validate_entries(struct bch_fs *c, struct jset *jset,
int write)
{
char buf[100];
struct jset_entry *entry;
int ret = 0;
vstruct_for_each(jset, entry) {
scnprintf(buf, sizeof(buf), "jset %llu entry offset %zi/%u",
le64_to_cpu(jset->seq),
(u64 *) entry - jset->_data,
le32_to_cpu(jset->u64s));
if (journal_entry_err_on(vstruct_next(entry) >
vstruct_last(jset), c,
"journal entry extends past end of jset")) {
jset->u64s = cpu_to_le32((u64 *) entry - jset->_data);
break;
}
ret = bch2_journal_entry_validate(c, buf, entry,
le32_to_cpu(jset->version),
JSET_BIG_ENDIAN(jset), write);
if (ret)
break;
}
fsck_err:
return ret;
}
static int jset_validate(struct bch_fs *c,
struct bch_dev *ca,
struct jset *jset, u64 sector,
unsigned bucket_sectors_left,
unsigned sectors_read,
int write)
{
size_t bytes = vstruct_bytes(jset);
struct bch_csum csum;
unsigned version;
int ret = 0;
if (le64_to_cpu(jset->magic) != jset_magic(c))
return JOURNAL_ENTRY_NONE;
version = le32_to_cpu(jset->version);
if (journal_entry_err_on((version != BCH_JSET_VERSION_OLD &&
version < bcachefs_metadata_version_min) ||
version >= bcachefs_metadata_version_max, c,
"%s sector %llu seq %llu: unknown journal entry version %u",
ca ? ca->name : c->name,
sector, le64_to_cpu(jset->seq),
version)) {
/* don't try to continue: */
return EINVAL;
}
if (bytes > (sectors_read << 9) &&
sectors_read < bucket_sectors_left)
return JOURNAL_ENTRY_REREAD;
if (journal_entry_err_on(bytes > bucket_sectors_left << 9, c,
"%s sector %llu seq %llu: journal entry too big (%zu bytes)",
ca ? ca->name : c->name,
sector, le64_to_cpu(jset->seq), bytes)) {
ret = JOURNAL_ENTRY_BAD;
le32_add_cpu(&jset->u64s,
-((bytes - (bucket_sectors_left << 9)) / 8));
}
if (journal_entry_err_on(!bch2_checksum_type_valid(c, JSET_CSUM_TYPE(jset)), c,
"%s sector %llu seq %llu: journal entry with unknown csum type %llu",
ca ? ca->name : c->name,
sector, le64_to_cpu(jset->seq),
JSET_CSUM_TYPE(jset))) {
ret = JOURNAL_ENTRY_BAD;
goto csum_done;
}
if (write)
goto csum_done;
csum = csum_vstruct(c, JSET_CSUM_TYPE(jset), journal_nonce(jset), jset);
if (journal_entry_err_on(bch2_crc_cmp(csum, jset->csum), c,
"%s sector %llu seq %llu: journal checksum bad",
ca ? ca->name : c->name,
sector, le64_to_cpu(jset->seq)))
ret = JOURNAL_ENTRY_BAD;
bch2_encrypt(c, JSET_CSUM_TYPE(jset), journal_nonce(jset),
jset->encrypted_start,
vstruct_end(jset) - (void *) jset->encrypted_start);
csum_done:
/* last_seq is ignored when JSET_NO_FLUSH is true */
if (journal_entry_err_on(!JSET_NO_FLUSH(jset) &&
le64_to_cpu(jset->last_seq) > le64_to_cpu(jset->seq), c,
"invalid journal entry: last_seq > seq (%llu > %llu)",
le64_to_cpu(jset->last_seq),
le64_to_cpu(jset->seq))) {
jset->last_seq = jset->seq;
return JOURNAL_ENTRY_BAD;
}
fsck_err:
return ret;
}
static int jset_validate_for_write(struct bch_fs *c, struct jset *jset)
{
unsigned sectors = vstruct_sectors(jset, c->block_bits);
return jset_validate(c, NULL, jset, 0, sectors, sectors, WRITE) ?:
jset_validate_entries(c, jset, WRITE);
}
struct journal_read_buf {
void *data;
size_t size;
};
static int journal_read_buf_realloc(struct journal_read_buf *b,
size_t new_size)
{
void *n;
/* the bios are sized for this many pages, max: */
if (new_size > JOURNAL_ENTRY_SIZE_MAX)
return -ENOMEM;
new_size = roundup_pow_of_two(new_size);
n = kvpmalloc(new_size, GFP_KERNEL);
if (!n)
return -ENOMEM;
kvpfree(b->data, b->size);
b->data = n;
b->size = new_size;
return 0;
}
static int journal_read_bucket(struct bch_dev *ca,
struct journal_read_buf *buf,
struct journal_list *jlist,
unsigned bucket)
{
struct bch_fs *c = ca->fs;
struct journal_device *ja = &ca->journal;
struct jset *j = NULL;
unsigned sectors, sectors_read = 0;
u64 offset = bucket_to_sector(ca, ja->buckets[bucket]),
end = offset + ca->mi.bucket_size;
bool saw_bad = false;
int ret = 0;
pr_debug("reading %u", bucket);
while (offset < end) {
if (!sectors_read) {
struct bio *bio;
unsigned nr_bvecs;
reread:
sectors_read = min_t(unsigned,
end - offset, buf->size >> 9);
nr_bvecs = buf_pages(buf->data, sectors_read << 9);
bio = bio_kmalloc(nr_bvecs, GFP_KERNEL);
bio_init(bio, ca->disk_sb.bdev, bio->bi_inline_vecs, nr_bvecs, REQ_OP_READ);
bio->bi_iter.bi_sector = offset;
bch2_bio_map(bio, buf->data, sectors_read << 9);
ret = submit_bio_wait(bio);
kfree(bio);
if (bch2_dev_io_err_on(ret, ca,
"journal read error: sector %llu",
offset) ||
bch2_meta_read_fault("journal")) {
/*
* We don't error out of the recovery process
* here, since the relevant journal entry may be
* found on a different device, and missing or
* no journal entries will be handled later
*/
return 0;
}
j = buf->data;
}
ret = jset_validate(c, ca, j, offset,
end - offset, sectors_read,
READ);
switch (ret) {
case BCH_FSCK_OK:
sectors = vstruct_sectors(j, c->block_bits);
break;
case JOURNAL_ENTRY_REREAD:
if (vstruct_bytes(j) > buf->size) {
ret = journal_read_buf_realloc(buf,
vstruct_bytes(j));
if (ret)
return ret;
}
goto reread;
case JOURNAL_ENTRY_NONE:
if (!saw_bad)
return 0;
sectors = block_sectors(c);
goto next_block;
case JOURNAL_ENTRY_BAD:
saw_bad = true;
/*
* On checksum error we don't really trust the size
* field of the journal entry we read, so try reading
* again at next block boundary:
*/
sectors = block_sectors(c);
break;
default:
return ret;
}
/*
* This happens sometimes if we don't have discards on -
* when we've partially overwritten a bucket with new
* journal entries. We don't need the rest of the
* bucket:
*/
if (le64_to_cpu(j->seq) < ja->bucket_seq[bucket])
return 0;
ja->bucket_seq[bucket] = le64_to_cpu(j->seq);
mutex_lock(&jlist->lock);
ret = journal_entry_add(c, ca, (struct bch_extent_ptr) {
.dev = ca->dev_idx,
.offset = offset,
}, jlist, j, ret != 0);
mutex_unlock(&jlist->lock);
switch (ret) {
case JOURNAL_ENTRY_ADD_OK:
break;
case JOURNAL_ENTRY_ADD_OUT_OF_RANGE:
break;
default:
return ret;
}
next_block:
pr_debug("next");
offset += sectors;
sectors_read -= sectors;
j = ((void *) j) + (sectors << 9);
}
return 0;
}
static void bch2_journal_read_device(struct closure *cl)
{
struct journal_device *ja =
container_of(cl, struct journal_device, read);
struct bch_dev *ca = container_of(ja, struct bch_dev, journal);
struct bch_fs *c = ca->fs;
struct journal_list *jlist =
container_of(cl->parent, struct journal_list, cl);
struct journal_read_buf buf = { NULL, 0 };
u64 min_seq = U64_MAX;
unsigned i;
int ret = 0;
if (!ja->nr)
goto out;
ret = journal_read_buf_realloc(&buf, PAGE_SIZE);
if (ret)
goto err;
pr_debug("%u journal buckets", ja->nr);
for (i = 0; i < ja->nr; i++) {
ret = journal_read_bucket(ca, &buf, jlist, i);
if (ret)
goto err;
}
/* Find the journal bucket with the highest sequence number: */
for (i = 0; i < ja->nr; i++) {
if (ja->bucket_seq[i] > ja->bucket_seq[ja->cur_idx])
ja->cur_idx = i;
min_seq = min(ja->bucket_seq[i], min_seq);
}
/*
* If there's duplicate journal entries in multiple buckets (which
* definitely isn't supposed to happen, but...) - make sure to start
* cur_idx at the last of those buckets, so we don't deadlock trying to
* allocate
*/
while (ja->bucket_seq[ja->cur_idx] > min_seq &&
ja->bucket_seq[ja->cur_idx] >
ja->bucket_seq[(ja->cur_idx + 1) % ja->nr])
ja->cur_idx = (ja->cur_idx + 1) % ja->nr;
ja->sectors_free = 0;
/*
* Set dirty_idx to indicate the entire journal is full and needs to be
* reclaimed - journal reclaim will immediately reclaim whatever isn't
* pinned when it first runs:
*/
ja->discard_idx = ja->dirty_idx_ondisk =
ja->dirty_idx = (ja->cur_idx + 1) % ja->nr;
out:
bch_verbose(c, "journal read done on device %s, ret %i", ca->name, ret);
kvpfree(buf.data, buf.size);
percpu_ref_put(&ca->io_ref);
closure_return(cl);
return;
err:
mutex_lock(&jlist->lock);
jlist->ret = ret;
mutex_unlock(&jlist->lock);
goto out;
}
static void bch2_journal_ptrs_to_text(struct printbuf *out, struct bch_fs *c,
struct journal_replay *j)
{
unsigned i;
for (i = 0; i < j->nr_ptrs; i++) {
struct bch_dev *ca = bch_dev_bkey_exists(c, j->ptrs[i].dev);
u64 offset;
div64_u64_rem(j->ptrs[i].offset, ca->mi.bucket_size, &offset);
if (i)
pr_buf(out, " ");
pr_buf(out, "%u:%llu (offset %llu)",
j->ptrs[i].dev,
(u64) j->ptrs[i].offset, offset);
}
}
bcachefs: Don't require flush/fua on every journal write This patch adds a flag to journal entries which, if set, indicates that they weren't done as flush/fua writes. - non flush/fua journal writes don't update last_seq (i.e. they don't free up space in the journal), thus the journal free space calculations now check whether nonflush journal writes are currently allowed (i.e. are we low on free space, or would doing a flush write free up a lot of space in the journal) - write_delay_ms, the user configurable option for when open journal entries are automatically written, is now interpreted as the max delay between flush journal writes (default 1 second). - bch2_journal_flush_seq_async is changed to ensure a flush write >= the requested sequence number has happened - journal read/replay must now ignore, and blacklist, any journal entries newer than the most recent flush entry in the journal. Also, the way the read_entire_journal option is handled has been improved; struct journal_replay now has an entry, 'ignore', for entries that were read but should not be used. - assorted refactoring and improvements related to journal read in journal_io.c and recovery.c Previously, we'd have to issue a flush/fua write every time we accumulated a full journal entry - typically the bucket size. Now we need to issue them much less frequently: when an fsync is requested, or it's been more than write_delay_ms since the last flush, or when we need to free up space in the journal. This is a significant performance improvement on many write heavy workloads. Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com> Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
2020-11-14 22:59:58 +08:00
int bch2_journal_read(struct bch_fs *c, struct list_head *list,
u64 *blacklist_seq, u64 *start_seq)
{
struct journal_list jlist;
bcachefs: Don't require flush/fua on every journal write This patch adds a flag to journal entries which, if set, indicates that they weren't done as flush/fua writes. - non flush/fua journal writes don't update last_seq (i.e. they don't free up space in the journal), thus the journal free space calculations now check whether nonflush journal writes are currently allowed (i.e. are we low on free space, or would doing a flush write free up a lot of space in the journal) - write_delay_ms, the user configurable option for when open journal entries are automatically written, is now interpreted as the max delay between flush journal writes (default 1 second). - bch2_journal_flush_seq_async is changed to ensure a flush write >= the requested sequence number has happened - journal read/replay must now ignore, and blacklist, any journal entries newer than the most recent flush entry in the journal. Also, the way the read_entire_journal option is handled has been improved; struct journal_replay now has an entry, 'ignore', for entries that were read but should not be used. - assorted refactoring and improvements related to journal read in journal_io.c and recovery.c Previously, we'd have to issue a flush/fua write every time we accumulated a full journal entry - typically the bucket size. Now we need to issue them much less frequently: when an fsync is requested, or it's been more than write_delay_ms since the last flush, or when we need to free up space in the journal. This is a significant performance improvement on many write heavy workloads. Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com> Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
2020-11-14 22:59:58 +08:00
struct journal_replay *i, *t;
struct bch_dev *ca;
unsigned iter;
size_t keys = 0, entries = 0;
bool degraded = false;
bcachefs: Don't require flush/fua on every journal write This patch adds a flag to journal entries which, if set, indicates that they weren't done as flush/fua writes. - non flush/fua journal writes don't update last_seq (i.e. they don't free up space in the journal), thus the journal free space calculations now check whether nonflush journal writes are currently allowed (i.e. are we low on free space, or would doing a flush write free up a lot of space in the journal) - write_delay_ms, the user configurable option for when open journal entries are automatically written, is now interpreted as the max delay between flush journal writes (default 1 second). - bch2_journal_flush_seq_async is changed to ensure a flush write >= the requested sequence number has happened - journal read/replay must now ignore, and blacklist, any journal entries newer than the most recent flush entry in the journal. Also, the way the read_entire_journal option is handled has been improved; struct journal_replay now has an entry, 'ignore', for entries that were read but should not be used. - assorted refactoring and improvements related to journal read in journal_io.c and recovery.c Previously, we'd have to issue a flush/fua write every time we accumulated a full journal entry - typically the bucket size. Now we need to issue them much less frequently: when an fsync is requested, or it's been more than write_delay_ms since the last flush, or when we need to free up space in the journal. This is a significant performance improvement on many write heavy workloads. Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com> Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
2020-11-14 22:59:58 +08:00
u64 seq, last_seq = 0;
int ret = 0;
closure_init_stack(&jlist.cl);
mutex_init(&jlist.lock);
jlist.head = list;
jlist.ret = 0;
for_each_member_device(ca, c, iter) {
if (!test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) &&
!(bch2_dev_has_data(c, ca) & (1 << BCH_DATA_journal)))
continue;
if ((ca->mi.state == BCH_MEMBER_STATE_rw ||
ca->mi.state == BCH_MEMBER_STATE_ro) &&
percpu_ref_tryget(&ca->io_ref))
closure_call(&ca->journal.read,
bch2_journal_read_device,
system_unbound_wq,
&jlist.cl);
else
degraded = true;
}
closure_sync(&jlist.cl);
if (jlist.ret)
return jlist.ret;
bcachefs: Don't require flush/fua on every journal write This patch adds a flag to journal entries which, if set, indicates that they weren't done as flush/fua writes. - non flush/fua journal writes don't update last_seq (i.e. they don't free up space in the journal), thus the journal free space calculations now check whether nonflush journal writes are currently allowed (i.e. are we low on free space, or would doing a flush write free up a lot of space in the journal) - write_delay_ms, the user configurable option for when open journal entries are automatically written, is now interpreted as the max delay between flush journal writes (default 1 second). - bch2_journal_flush_seq_async is changed to ensure a flush write >= the requested sequence number has happened - journal read/replay must now ignore, and blacklist, any journal entries newer than the most recent flush entry in the journal. Also, the way the read_entire_journal option is handled has been improved; struct journal_replay now has an entry, 'ignore', for entries that were read but should not be used. - assorted refactoring and improvements related to journal read in journal_io.c and recovery.c Previously, we'd have to issue a flush/fua write every time we accumulated a full journal entry - typically the bucket size. Now we need to issue them much less frequently: when an fsync is requested, or it's been more than write_delay_ms since the last flush, or when we need to free up space in the journal. This is a significant performance improvement on many write heavy workloads. Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com> Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
2020-11-14 22:59:58 +08:00
if (list_empty(list)) {
bch_info(c, "journal read done, but no entries found");
return 0;
}
i = list_last_entry(list, struct journal_replay, list);
*start_seq = le64_to_cpu(i->j.seq) + 1;
/*
* Find most recent flush entry, and ignore newer non flush entries -
* those entries will be blacklisted:
*/
list_for_each_entry_safe_reverse(i, t, list, list) {
if (i->ignore)
continue;
if (!JSET_NO_FLUSH(&i->j)) {
last_seq = le64_to_cpu(i->j.last_seq);
*blacklist_seq = le64_to_cpu(i->j.seq) + 1;
break;
}
journal_replay_free(c, i);
}
if (!last_seq) {
fsck_err(c, "journal read done, but no entries found after dropping non-flushes");
return -1;
}
/* Drop blacklisted entries and entries older than last_seq: */
list_for_each_entry_safe(i, t, list, list) {
if (i->ignore)
continue;
seq = le64_to_cpu(i->j.seq);
if (seq < last_seq) {
journal_replay_free(c, i);
continue;
}
if (bch2_journal_seq_is_blacklisted(c, seq, true)) {
fsck_err_on(!JSET_NO_FLUSH(&i->j), c,
"found blacklisted journal entry %llu", seq);
journal_replay_free(c, i);
}
}
/* Check for missing entries: */
seq = last_seq;
list_for_each_entry(i, list, list) {
if (i->ignore)
continue;
BUG_ON(seq > le64_to_cpu(i->j.seq));
while (seq < le64_to_cpu(i->j.seq)) {
u64 missing_start, missing_end;
char buf1[200], buf2[200];
bcachefs: Don't require flush/fua on every journal write This patch adds a flag to journal entries which, if set, indicates that they weren't done as flush/fua writes. - non flush/fua journal writes don't update last_seq (i.e. they don't free up space in the journal), thus the journal free space calculations now check whether nonflush journal writes are currently allowed (i.e. are we low on free space, or would doing a flush write free up a lot of space in the journal) - write_delay_ms, the user configurable option for when open journal entries are automatically written, is now interpreted as the max delay between flush journal writes (default 1 second). - bch2_journal_flush_seq_async is changed to ensure a flush write >= the requested sequence number has happened - journal read/replay must now ignore, and blacklist, any journal entries newer than the most recent flush entry in the journal. Also, the way the read_entire_journal option is handled has been improved; struct journal_replay now has an entry, 'ignore', for entries that were read but should not be used. - assorted refactoring and improvements related to journal read in journal_io.c and recovery.c Previously, we'd have to issue a flush/fua write every time we accumulated a full journal entry - typically the bucket size. Now we need to issue them much less frequently: when an fsync is requested, or it's been more than write_delay_ms since the last flush, or when we need to free up space in the journal. This is a significant performance improvement on many write heavy workloads. Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com> Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
2020-11-14 22:59:58 +08:00
while (seq < le64_to_cpu(i->j.seq) &&
bch2_journal_seq_is_blacklisted(c, seq, false))
seq++;
if (seq == le64_to_cpu(i->j.seq))
break;
missing_start = seq;
while (seq < le64_to_cpu(i->j.seq) &&
!bch2_journal_seq_is_blacklisted(c, seq, false))
seq++;
if (i->list.prev != list) {
struct printbuf out = PBUF(buf1);
struct journal_replay *p = list_prev_entry(i, list);
bch2_journal_ptrs_to_text(&out, c, p);
pr_buf(&out, " size %llu", vstruct_sectors(&p->j, c->block_bits));
} else
sprintf(buf1, "(none)");
bch2_journal_ptrs_to_text(&PBUF(buf2), c, i);
bcachefs: Don't require flush/fua on every journal write This patch adds a flag to journal entries which, if set, indicates that they weren't done as flush/fua writes. - non flush/fua journal writes don't update last_seq (i.e. they don't free up space in the journal), thus the journal free space calculations now check whether nonflush journal writes are currently allowed (i.e. are we low on free space, or would doing a flush write free up a lot of space in the journal) - write_delay_ms, the user configurable option for when open journal entries are automatically written, is now interpreted as the max delay between flush journal writes (default 1 second). - bch2_journal_flush_seq_async is changed to ensure a flush write >= the requested sequence number has happened - journal read/replay must now ignore, and blacklist, any journal entries newer than the most recent flush entry in the journal. Also, the way the read_entire_journal option is handled has been improved; struct journal_replay now has an entry, 'ignore', for entries that were read but should not be used. - assorted refactoring and improvements related to journal read in journal_io.c and recovery.c Previously, we'd have to issue a flush/fua write every time we accumulated a full journal entry - typically the bucket size. Now we need to issue them much less frequently: when an fsync is requested, or it's been more than write_delay_ms since the last flush, or when we need to free up space in the journal. This is a significant performance improvement on many write heavy workloads. Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com> Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
2020-11-14 22:59:58 +08:00
missing_end = seq - 1;
fsck_err(c, "journal entries %llu-%llu missing! (replaying %llu-%llu)\n"
" prev at %s\n"
" next at %s",
bcachefs: Don't require flush/fua on every journal write This patch adds a flag to journal entries which, if set, indicates that they weren't done as flush/fua writes. - non flush/fua journal writes don't update last_seq (i.e. they don't free up space in the journal), thus the journal free space calculations now check whether nonflush journal writes are currently allowed (i.e. are we low on free space, or would doing a flush write free up a lot of space in the journal) - write_delay_ms, the user configurable option for when open journal entries are automatically written, is now interpreted as the max delay between flush journal writes (default 1 second). - bch2_journal_flush_seq_async is changed to ensure a flush write >= the requested sequence number has happened - journal read/replay must now ignore, and blacklist, any journal entries newer than the most recent flush entry in the journal. Also, the way the read_entire_journal option is handled has been improved; struct journal_replay now has an entry, 'ignore', for entries that were read but should not be used. - assorted refactoring and improvements related to journal read in journal_io.c and recovery.c Previously, we'd have to issue a flush/fua write every time we accumulated a full journal entry - typically the bucket size. Now we need to issue them much less frequently: when an fsync is requested, or it's been more than write_delay_ms since the last flush, or when we need to free up space in the journal. This is a significant performance improvement on many write heavy workloads. Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com> Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
2020-11-14 22:59:58 +08:00
missing_start, missing_end,
last_seq, *blacklist_seq - 1,
buf1, buf2);
bcachefs: Don't require flush/fua on every journal write This patch adds a flag to journal entries which, if set, indicates that they weren't done as flush/fua writes. - non flush/fua journal writes don't update last_seq (i.e. they don't free up space in the journal), thus the journal free space calculations now check whether nonflush journal writes are currently allowed (i.e. are we low on free space, or would doing a flush write free up a lot of space in the journal) - write_delay_ms, the user configurable option for when open journal entries are automatically written, is now interpreted as the max delay between flush journal writes (default 1 second). - bch2_journal_flush_seq_async is changed to ensure a flush write >= the requested sequence number has happened - journal read/replay must now ignore, and blacklist, any journal entries newer than the most recent flush entry in the journal. Also, the way the read_entire_journal option is handled has been improved; struct journal_replay now has an entry, 'ignore', for entries that were read but should not be used. - assorted refactoring and improvements related to journal read in journal_io.c and recovery.c Previously, we'd have to issue a flush/fua write every time we accumulated a full journal entry - typically the bucket size. Now we need to issue them much less frequently: when an fsync is requested, or it's been more than write_delay_ms since the last flush, or when we need to free up space in the journal. This is a significant performance improvement on many write heavy workloads. Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com> Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
2020-11-14 22:59:58 +08:00
}
seq++;
}
list_for_each_entry(i, list, list) {
struct jset_entry *entry;
struct bkey_i *k, *_n;
struct bch_replicas_padded replicas = {
.e.data_type = BCH_DATA_journal,
.e.nr_required = 1,
};
unsigned ptr;
char buf[80];
bcachefs: Don't require flush/fua on every journal write This patch adds a flag to journal entries which, if set, indicates that they weren't done as flush/fua writes. - non flush/fua journal writes don't update last_seq (i.e. they don't free up space in the journal), thus the journal free space calculations now check whether nonflush journal writes are currently allowed (i.e. are we low on free space, or would doing a flush write free up a lot of space in the journal) - write_delay_ms, the user configurable option for when open journal entries are automatically written, is now interpreted as the max delay between flush journal writes (default 1 second). - bch2_journal_flush_seq_async is changed to ensure a flush write >= the requested sequence number has happened - journal read/replay must now ignore, and blacklist, any journal entries newer than the most recent flush entry in the journal. Also, the way the read_entire_journal option is handled has been improved; struct journal_replay now has an entry, 'ignore', for entries that were read but should not be used. - assorted refactoring and improvements related to journal read in journal_io.c and recovery.c Previously, we'd have to issue a flush/fua write every time we accumulated a full journal entry - typically the bucket size. Now we need to issue them much less frequently: when an fsync is requested, or it's been more than write_delay_ms since the last flush, or when we need to free up space in the journal. This is a significant performance improvement on many write heavy workloads. Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com> Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
2020-11-14 22:59:58 +08:00
if (i->ignore)
continue;
ret = jset_validate_entries(c, &i->j, READ);
if (ret)
goto fsck_err;
for (ptr = 0; ptr < i->nr_ptrs; ptr++)
replicas.e.devs[replicas.e.nr_devs++] = i->ptrs[ptr].dev;
bch2_replicas_entry_sort(&replicas.e);
/*
* If we're mounting in degraded mode - if we didn't read all
* the devices - this is wrong:
*/
if (!degraded &&
(test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) ||
fsck_err_on(!bch2_replicas_marked(c, &replicas.e), c,
"superblock not marked as containing replicas %s",
(bch2_replicas_entry_to_text(&PBUF(buf),
&replicas.e), buf)))) {
ret = bch2_mark_replicas(c, &replicas.e);
if (ret)
return ret;
}
for_each_jset_key(k, _n, entry, &i->j)
keys++;
entries++;
}
bcachefs: Don't require flush/fua on every journal write This patch adds a flag to journal entries which, if set, indicates that they weren't done as flush/fua writes. - non flush/fua journal writes don't update last_seq (i.e. they don't free up space in the journal), thus the journal free space calculations now check whether nonflush journal writes are currently allowed (i.e. are we low on free space, or would doing a flush write free up a lot of space in the journal) - write_delay_ms, the user configurable option for when open journal entries are automatically written, is now interpreted as the max delay between flush journal writes (default 1 second). - bch2_journal_flush_seq_async is changed to ensure a flush write >= the requested sequence number has happened - journal read/replay must now ignore, and blacklist, any journal entries newer than the most recent flush entry in the journal. Also, the way the read_entire_journal option is handled has been improved; struct journal_replay now has an entry, 'ignore', for entries that were read but should not be used. - assorted refactoring and improvements related to journal read in journal_io.c and recovery.c Previously, we'd have to issue a flush/fua write every time we accumulated a full journal entry - typically the bucket size. Now we need to issue them much less frequently: when an fsync is requested, or it's been more than write_delay_ms since the last flush, or when we need to free up space in the journal. This is a significant performance improvement on many write heavy workloads. Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com> Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
2020-11-14 22:59:58 +08:00
bch_info(c, "journal read done, %zu keys in %zu entries, seq %llu",
keys, entries, *start_seq);
bcachefs: Don't require flush/fua on every journal write This patch adds a flag to journal entries which, if set, indicates that they weren't done as flush/fua writes. - non flush/fua journal writes don't update last_seq (i.e. they don't free up space in the journal), thus the journal free space calculations now check whether nonflush journal writes are currently allowed (i.e. are we low on free space, or would doing a flush write free up a lot of space in the journal) - write_delay_ms, the user configurable option for when open journal entries are automatically written, is now interpreted as the max delay between flush journal writes (default 1 second). - bch2_journal_flush_seq_async is changed to ensure a flush write >= the requested sequence number has happened - journal read/replay must now ignore, and blacklist, any journal entries newer than the most recent flush entry in the journal. Also, the way the read_entire_journal option is handled has been improved; struct journal_replay now has an entry, 'ignore', for entries that were read but should not be used. - assorted refactoring and improvements related to journal read in journal_io.c and recovery.c Previously, we'd have to issue a flush/fua write every time we accumulated a full journal entry - typically the bucket size. Now we need to issue them much less frequently: when an fsync is requested, or it's been more than write_delay_ms since the last flush, or when we need to free up space in the journal. This is a significant performance improvement on many write heavy workloads. Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com> Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
2020-11-14 22:59:58 +08:00
if (*start_seq != *blacklist_seq)
bch_info(c, "dropped unflushed entries %llu-%llu",
*blacklist_seq, *start_seq - 1);
fsck_err:
return ret;
}
/* journal write: */
static void __journal_write_alloc(struct journal *j,
struct journal_buf *w,
struct dev_alloc_list *devs_sorted,
unsigned sectors,
unsigned *replicas,
unsigned replicas_want)
{
struct bch_fs *c = container_of(j, struct bch_fs, journal);
struct journal_device *ja;
struct bch_dev *ca;
unsigned i;
if (*replicas >= replicas_want)
return;
for (i = 0; i < devs_sorted->nr; i++) {
ca = rcu_dereference(c->devs[devs_sorted->devs[i]]);
if (!ca)
continue;
ja = &ca->journal;
/*
* Check that we can use this device, and aren't already using
* it:
*/
if (!ca->mi.durability ||
ca->mi.state != BCH_MEMBER_STATE_rw ||
!ja->nr ||
bch2_bkey_has_device(bkey_i_to_s_c(&w->key),
ca->dev_idx) ||
sectors > ja->sectors_free)
continue;
bch2_dev_stripe_increment(ca, &j->wp.stripe);
bch2_bkey_append_ptr(&w->key,
(struct bch_extent_ptr) {
.offset = bucket_to_sector(ca,
ja->buckets[ja->cur_idx]) +
ca->mi.bucket_size -
ja->sectors_free,
.dev = ca->dev_idx,
});
ja->sectors_free -= sectors;
ja->bucket_seq[ja->cur_idx] = le64_to_cpu(w->data->seq);
*replicas += ca->mi.durability;
if (*replicas >= replicas_want)
break;
}
}
/**
* journal_next_bucket - move on to the next journal bucket if possible
*/
static int journal_write_alloc(struct journal *j, struct journal_buf *w,
unsigned sectors)
{
struct bch_fs *c = container_of(j, struct bch_fs, journal);
struct bch_devs_mask devs;
struct journal_device *ja;
struct bch_dev *ca;
struct dev_alloc_list devs_sorted;
unsigned target = c->opts.metadata_target ?:
c->opts.foreground_target;
unsigned i, replicas = 0, replicas_want =
READ_ONCE(c->opts.metadata_replicas);
rcu_read_lock();
retry:
devs = target_rw_devs(c, BCH_DATA_journal, target);
devs_sorted = bch2_dev_alloc_list(c, &j->wp.stripe, &devs);
__journal_write_alloc(j, w, &devs_sorted,
sectors, &replicas, replicas_want);
if (replicas >= replicas_want)
goto done;
for (i = 0; i < devs_sorted.nr; i++) {
ca = rcu_dereference(c->devs[devs_sorted.devs[i]]);
if (!ca)
continue;
ja = &ca->journal;
if (sectors > ja->sectors_free &&
sectors <= ca->mi.bucket_size &&
bch2_journal_dev_buckets_available(j, ja,
journal_space_discarded)) {
ja->cur_idx = (ja->cur_idx + 1) % ja->nr;
ja->sectors_free = ca->mi.bucket_size;
/*
* ja->bucket_seq[ja->cur_idx] must always have
* something sensible:
*/
ja->bucket_seq[ja->cur_idx] = le64_to_cpu(w->data->seq);
}
}
__journal_write_alloc(j, w, &devs_sorted,
sectors, &replicas, replicas_want);
if (replicas < replicas_want && target) {
/* Retry from all devices: */
target = 0;
goto retry;
}
done:
rcu_read_unlock();
BUG_ON(bkey_val_u64s(&w->key.k) > BCH_REPLICAS_MAX);
return replicas >= c->opts.metadata_replicas_required ? 0 : -EROFS;
}
static void journal_write_compact(struct jset *jset)
{
struct jset_entry *i, *next, *prev = NULL;
/*
* Simple compaction, dropping empty jset_entries (from journal
* reservations that weren't fully used) and merging jset_entries that
* can be.
*
* If we wanted to be really fancy here, we could sort all the keys in
* the jset and drop keys that were overwritten - probably not worth it:
*/
vstruct_for_each_safe(jset, i, next) {
unsigned u64s = le16_to_cpu(i->u64s);
/* Empty entry: */
if (!u64s)
continue;
/* Can we merge with previous entry? */
if (prev &&
i->btree_id == prev->btree_id &&
i->level == prev->level &&
i->type == prev->type &&
i->type == BCH_JSET_ENTRY_btree_keys &&
le16_to_cpu(prev->u64s) + u64s <= U16_MAX) {
memmove_u64s_down(vstruct_next(prev),
i->_data,
u64s);
le16_add_cpu(&prev->u64s, u64s);
continue;
}
/* Couldn't merge, move i into new position (after prev): */
prev = prev ? vstruct_next(prev) : jset->start;
if (i != prev)
memmove_u64s_down(prev, i, jset_u64s(u64s));
}
prev = prev ? vstruct_next(prev) : jset->start;
jset->u64s = cpu_to_le32((u64 *) prev - jset->_data);
}
static void journal_buf_realloc(struct journal *j, struct journal_buf *buf)
{
/* we aren't holding j->lock: */
unsigned new_size = READ_ONCE(j->buf_size_want);
void *new_buf;
if (buf->buf_size >= new_size)
return;
new_buf = kvpmalloc(new_size, GFP_NOIO|__GFP_NOWARN);
if (!new_buf)
return;
memcpy(new_buf, buf->data, buf->buf_size);
spin_lock(&j->lock);
swap(buf->data, new_buf);
swap(buf->buf_size, new_size);
spin_unlock(&j->lock);
kvpfree(new_buf, new_size);
}
static inline struct journal_buf *journal_last_unwritten_buf(struct journal *j)
{
return j->buf + j->reservations.unwritten_idx;
}
static void journal_write_done(struct closure *cl)
{
struct journal *j = container_of(cl, struct journal, io);
struct bch_fs *c = container_of(j, struct bch_fs, journal);
struct journal_buf *w = journal_last_unwritten_buf(j);
struct bch_replicas_padded replicas;
union journal_res_state old, new;
u64 v, seq;
int err = 0;
bch2_time_stats_update(!JSET_NO_FLUSH(w->data)
? j->flush_write_time
: j->noflush_write_time, j->write_start_time);
if (!w->devs_written.nr) {
bch_err(c, "unable to write journal to sufficient devices");
err = -EIO;
} else {
bch2_devlist_to_replicas(&replicas.e, BCH_DATA_journal,
w->devs_written);
if (bch2_mark_replicas(c, &replicas.e))
err = -EIO;
}
if (err)
bch2_fatal_error(c);
spin_lock(&j->lock);
seq = le64_to_cpu(w->data->seq);
if (seq >= j->pin.front)
journal_seq_pin(j, seq)->devs = w->devs_written;
if (!err) {
j->seq_ondisk = seq;
bcachefs: Don't require flush/fua on every journal write This patch adds a flag to journal entries which, if set, indicates that they weren't done as flush/fua writes. - non flush/fua journal writes don't update last_seq (i.e. they don't free up space in the journal), thus the journal free space calculations now check whether nonflush journal writes are currently allowed (i.e. are we low on free space, or would doing a flush write free up a lot of space in the journal) - write_delay_ms, the user configurable option for when open journal entries are automatically written, is now interpreted as the max delay between flush journal writes (default 1 second). - bch2_journal_flush_seq_async is changed to ensure a flush write >= the requested sequence number has happened - journal read/replay must now ignore, and blacklist, any journal entries newer than the most recent flush entry in the journal. Also, the way the read_entire_journal option is handled has been improved; struct journal_replay now has an entry, 'ignore', for entries that were read but should not be used. - assorted refactoring and improvements related to journal read in journal_io.c and recovery.c Previously, we'd have to issue a flush/fua write every time we accumulated a full journal entry - typically the bucket size. Now we need to issue them much less frequently: when an fsync is requested, or it's been more than write_delay_ms since the last flush, or when we need to free up space in the journal. This is a significant performance improvement on many write heavy workloads. Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com> Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
2020-11-14 22:59:58 +08:00
if (!JSET_NO_FLUSH(w->data)) {
j->flushed_seq_ondisk = seq;
j->last_seq_ondisk = w->last_seq;
}
} else if (!j->err_seq || seq < j->err_seq)
j->err_seq = seq;
/*
* Updating last_seq_ondisk may let bch2_journal_reclaim_work() discard
* more buckets:
*
* Must come before signaling write completion, for
* bch2_fs_journal_stop():
*/
journal_reclaim_kick(&c->journal);
/* also must come before signalling write completion: */
closure_debug_destroy(cl);
v = atomic64_read(&j->reservations.counter);
do {
old.v = new.v = v;
BUG_ON(new.idx == new.unwritten_idx);
new.unwritten_idx++;
} while ((v = atomic64_cmpxchg(&j->reservations.counter,
old.v, new.v)) != old.v);
bch2_journal_space_available(j);
closure_wake_up(&w->wait);
journal_wake(j);
if (test_bit(JOURNAL_NEED_WRITE, &j->flags))
mod_delayed_work(c->io_complete_wq, &j->write_work, 0);
spin_unlock(&j->lock);
if (new.unwritten_idx != new.idx &&
!journal_state_count(new, new.unwritten_idx))
closure_call(&j->io, bch2_journal_write, c->io_complete_wq, NULL);
}
static void journal_write_endio(struct bio *bio)
{
struct bch_dev *ca = bio->bi_private;
struct journal *j = &ca->fs->journal;
struct journal_buf *w = journal_last_unwritten_buf(j);
unsigned long flags;
if (bch2_dev_io_err_on(bio->bi_status, ca, "error writing journal entry %llu: %s",
le64_to_cpu(w->data->seq),
bch2_blk_status_to_str(bio->bi_status)) ||
bch2_meta_write_fault("journal")) {
spin_lock_irqsave(&j->err_lock, flags);
bch2_dev_list_drop_dev(&w->devs_written, ca->dev_idx);
spin_unlock_irqrestore(&j->err_lock, flags);
}
closure_put(&j->io);
percpu_ref_put(&ca->io_ref);
}
static void do_journal_write(struct closure *cl)
{
struct journal *j = container_of(cl, struct journal, io);
struct bch_fs *c = container_of(j, struct bch_fs, journal);
struct bch_dev *ca;
struct journal_buf *w = journal_last_unwritten_buf(j);
struct bch_extent_ptr *ptr;
struct bio *bio;
unsigned sectors = vstruct_sectors(w->data, c->block_bits);
extent_for_each_ptr(bkey_i_to_s_extent(&w->key), ptr) {
ca = bch_dev_bkey_exists(c, ptr->dev);
if (!percpu_ref_tryget(&ca->io_ref)) {
/* XXX: fix this */
bch_err(c, "missing device for journal write\n");
continue;
}
this_cpu_add(ca->io_done->sectors[WRITE][BCH_DATA_journal],
sectors);
bio = ca->journal.bio;
bio_reset(bio, ca->disk_sb.bdev, REQ_OP_WRITE|REQ_SYNC|REQ_META);
bio->bi_iter.bi_sector = ptr->offset;
bio->bi_end_io = journal_write_endio;
bio->bi_private = ca;
BUG_ON(bio->bi_iter.bi_sector == ca->prev_journal_sector);
ca->prev_journal_sector = bio->bi_iter.bi_sector;
if (!JSET_NO_FLUSH(w->data))
bio->bi_opf |= REQ_FUA;
if (!JSET_NO_FLUSH(w->data) && !w->separate_flush)
bio->bi_opf |= REQ_PREFLUSH;
bch2_bio_map(bio, w->data, sectors << 9);
trace_journal_write(bio);
closure_bio_submit(bio, cl);
ca->journal.bucket_seq[ca->journal.cur_idx] =
le64_to_cpu(w->data->seq);
}
continue_at(cl, journal_write_done, c->io_complete_wq);
return;
}
void bch2_journal_write(struct closure *cl)
{
struct journal *j = container_of(cl, struct journal, io);
struct bch_fs *c = container_of(j, struct bch_fs, journal);
struct bch_dev *ca;
struct journal_buf *w = journal_last_unwritten_buf(j);
struct jset_entry *start, *end;
struct jset *jset;
struct bio *bio;
char *journal_debug_buf = NULL;
bool validate_before_checksum = false;
unsigned i, sectors, bytes, u64s, nr_rw_members = 0;
int ret;
BUG_ON(BCH_SB_CLEAN(c->disk_sb.sb));
journal_buf_realloc(j, w);
jset = w->data;
j->write_start_time = local_clock();
bcachefs: Don't require flush/fua on every journal write This patch adds a flag to journal entries which, if set, indicates that they weren't done as flush/fua writes. - non flush/fua journal writes don't update last_seq (i.e. they don't free up space in the journal), thus the journal free space calculations now check whether nonflush journal writes are currently allowed (i.e. are we low on free space, or would doing a flush write free up a lot of space in the journal) - write_delay_ms, the user configurable option for when open journal entries are automatically written, is now interpreted as the max delay between flush journal writes (default 1 second). - bch2_journal_flush_seq_async is changed to ensure a flush write >= the requested sequence number has happened - journal read/replay must now ignore, and blacklist, any journal entries newer than the most recent flush entry in the journal. Also, the way the read_entire_journal option is handled has been improved; struct journal_replay now has an entry, 'ignore', for entries that were read but should not be used. - assorted refactoring and improvements related to journal read in journal_io.c and recovery.c Previously, we'd have to issue a flush/fua write every time we accumulated a full journal entry - typically the bucket size. Now we need to issue them much less frequently: when an fsync is requested, or it's been more than write_delay_ms since the last flush, or when we need to free up space in the journal. This is a significant performance improvement on many write heavy workloads. Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com> Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
2020-11-14 22:59:58 +08:00
spin_lock(&j->lock);
if (c->sb.features & (1ULL << BCH_FEATURE_journal_no_flush) &&
(w->noflush ||
(!w->must_flush &&
(jiffies - j->last_flush_write) < msecs_to_jiffies(c->opts.journal_flush_delay) &&
test_bit(JOURNAL_MAY_SKIP_FLUSH, &j->flags)))) {
bcachefs: Don't require flush/fua on every journal write This patch adds a flag to journal entries which, if set, indicates that they weren't done as flush/fua writes. - non flush/fua journal writes don't update last_seq (i.e. they don't free up space in the journal), thus the journal free space calculations now check whether nonflush journal writes are currently allowed (i.e. are we low on free space, or would doing a flush write free up a lot of space in the journal) - write_delay_ms, the user configurable option for when open journal entries are automatically written, is now interpreted as the max delay between flush journal writes (default 1 second). - bch2_journal_flush_seq_async is changed to ensure a flush write >= the requested sequence number has happened - journal read/replay must now ignore, and blacklist, any journal entries newer than the most recent flush entry in the journal. Also, the way the read_entire_journal option is handled has been improved; struct journal_replay now has an entry, 'ignore', for entries that were read but should not be used. - assorted refactoring and improvements related to journal read in journal_io.c and recovery.c Previously, we'd have to issue a flush/fua write every time we accumulated a full journal entry - typically the bucket size. Now we need to issue them much less frequently: when an fsync is requested, or it's been more than write_delay_ms since the last flush, or when we need to free up space in the journal. This is a significant performance improvement on many write heavy workloads. Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com> Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
2020-11-14 22:59:58 +08:00
w->noflush = true;
SET_JSET_NO_FLUSH(jset, true);
jset->last_seq = 0;
w->last_seq = 0;
bcachefs: Don't require flush/fua on every journal write This patch adds a flag to journal entries which, if set, indicates that they weren't done as flush/fua writes. - non flush/fua journal writes don't update last_seq (i.e. they don't free up space in the journal), thus the journal free space calculations now check whether nonflush journal writes are currently allowed (i.e. are we low on free space, or would doing a flush write free up a lot of space in the journal) - write_delay_ms, the user configurable option for when open journal entries are automatically written, is now interpreted as the max delay between flush journal writes (default 1 second). - bch2_journal_flush_seq_async is changed to ensure a flush write >= the requested sequence number has happened - journal read/replay must now ignore, and blacklist, any journal entries newer than the most recent flush entry in the journal. Also, the way the read_entire_journal option is handled has been improved; struct journal_replay now has an entry, 'ignore', for entries that were read but should not be used. - assorted refactoring and improvements related to journal read in journal_io.c and recovery.c Previously, we'd have to issue a flush/fua write every time we accumulated a full journal entry - typically the bucket size. Now we need to issue them much less frequently: when an fsync is requested, or it's been more than write_delay_ms since the last flush, or when we need to free up space in the journal. This is a significant performance improvement on many write heavy workloads. Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com> Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
2020-11-14 22:59:58 +08:00
j->nr_noflush_writes++;
} else {
j->last_flush_write = jiffies;
j->nr_flush_writes++;
}
spin_unlock(&j->lock);
/*
* New btree roots are set by journalling them; when the journal entry
* gets written we have to propagate them to c->btree_roots
*
* But, every journal entry we write has to contain all the btree roots
* (at least for now); so after we copy btree roots to c->btree_roots we
* have to get any missing btree roots and add them to this journal
* entry:
*/
bch2_journal_entries_to_btree_roots(c, jset);
start = end = vstruct_last(jset);
end = bch2_btree_roots_to_journal_entries(c, jset->start, end);
bch2_journal_super_entries_add_common(c, &end,
le64_to_cpu(jset->seq));
u64s = (u64 *) end - (u64 *) start;
BUG_ON(u64s > j->entry_u64s_reserved);
le32_add_cpu(&jset->u64s, u64s);
BUG_ON(vstruct_sectors(jset, c->block_bits) > w->sectors);
journal_write_compact(jset);
jset->magic = cpu_to_le64(jset_magic(c));
jset->version = c->sb.version < bcachefs_metadata_version_new_versioning
? cpu_to_le32(BCH_JSET_VERSION_OLD)
: cpu_to_le32(c->sb.version);
SET_JSET_BIG_ENDIAN(jset, CPU_BIG_ENDIAN);
SET_JSET_CSUM_TYPE(jset, bch2_meta_checksum_type(c));
if (!JSET_NO_FLUSH(jset) && journal_entry_empty(jset))
j->last_empty_seq = le64_to_cpu(jset->seq);
if (bch2_csum_type_is_encryption(JSET_CSUM_TYPE(jset)))
validate_before_checksum = true;
if (le32_to_cpu(jset->version) < bcachefs_metadata_version_current)
validate_before_checksum = true;
if (validate_before_checksum &&
jset_validate_for_write(c, jset))
goto err;
bch2_encrypt(c, JSET_CSUM_TYPE(jset), journal_nonce(jset),
jset->encrypted_start,
vstruct_end(jset) - (void *) jset->encrypted_start);
jset->csum = csum_vstruct(c, JSET_CSUM_TYPE(jset),
journal_nonce(jset), jset);
if (!validate_before_checksum &&
jset_validate_for_write(c, jset))
goto err;
sectors = vstruct_sectors(jset, c->block_bits);
BUG_ON(sectors > w->sectors);
bytes = vstruct_bytes(jset);
memset((void *) jset + bytes, 0, (sectors << 9) - bytes);
retry_alloc:
spin_lock(&j->lock);
ret = journal_write_alloc(j, w, sectors);
if (ret && j->can_discard) {
spin_unlock(&j->lock);
bch2_journal_do_discards(j);
goto retry_alloc;
}
if (ret) {
journal_debug_buf = kmalloc(4096, GFP_ATOMIC);
if (journal_debug_buf)
__bch2_journal_debug_to_text(&_PBUF(journal_debug_buf, 4096), j);
}
/*
* write is allocated, no longer need to account for it in
* bch2_journal_space_available():
*/
w->sectors = 0;
/*
* journal entry has been compacted and allocated, recalculate space
* available:
*/
bch2_journal_space_available(j);
spin_unlock(&j->lock);
if (ret) {
bch_err(c, "Unable to allocate journal write:\n%s",
journal_debug_buf);
kfree(journal_debug_buf);
bch2_fatal_error(c);
continue_at(cl, journal_write_done, c->io_complete_wq);
return;
}
w->devs_written = bch2_bkey_devs(bkey_i_to_s_c(&w->key));
if (test_bit(JOURNAL_NOCHANGES, &j->flags))
goto no_io;
for_each_rw_member(ca, c, i)
nr_rw_members++;
if (nr_rw_members > 1)
w->separate_flush = true;
if (!JSET_NO_FLUSH(jset) && w->separate_flush) {
for_each_rw_member(ca, c, i) {
percpu_ref_get(&ca->io_ref);
bio = ca->journal.bio;
bio_reset(bio, ca->disk_sb.bdev, REQ_OP_FLUSH);
bio->bi_end_io = journal_write_endio;
bio->bi_private = ca;
closure_bio_submit(bio, cl);
}
}
continue_at(cl, do_journal_write, c->io_complete_wq);
return;
no_io:
continue_at(cl, journal_write_done, c->io_complete_wq);
return;
err:
bch2_fatal_error(c);
continue_at(cl, journal_write_done, c->io_complete_wq);
}