unimore/kernel-hacking-2024-linux-steffo
unimore/kernel-hacking-2024-linux-steffo
1
Fork 0
Code Pull requests 5 Wiki Activity

bcachefs: Refactor open_bucket code

Browse source 
Prep work for adding a hash table of open buckets - instead of embedding
a bch_extent_ptr, we need to refer to the bucket directly so that we're
not calling sector_to_bucket() in the hash table lookup code, which has
an expensive divide.

Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
This commit is contained in:
Kent Overstreet 2021-12-25 21:21:46 -05:00 committed by Kent Overstreet
parent 57af63b286
commit abe19d458e
8 changed files with 83 additions and 67 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

21
fs/bcachefs/alloc_background.c
View file

@ -1066,7 +1066,7 @@ static bool bch2_dev_has_open_write_point(struct bch_fs *c, struct bch_dev *ca)
ob++) {
spin_lock(&ob->lock);
if (ob->valid && !ob->on_partial_list &&
ob->ptr.dev == ca->dev_idx)
ob->dev == ca->dev_idx)
ret = true;
spin_unlock(&ob->lock);
}
@ -1213,22 +1213,3 @@ void bch2_fs_allocator_background_init(struct bch_fs *c)
{
spin_lock_init(&c->freelist_lock);
}
void bch2_open_buckets_to_text(struct printbuf *out, struct bch_fs *c)
{
struct open_bucket *ob;
for (ob = c->open_buckets;
ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
ob++) {
spin_lock(&ob->lock);
if (ob->valid && !ob->on_partial_list) {
pr_buf(out, "%zu ref %u type %s\n",
ob - c->open_buckets,
atomic_read(&ob->pin),
bch2_data_types[ob->type]);
}
spin_unlock(&ob->lock);
}
}

2
fs/bcachefs/alloc_background.h
View file

@ -142,6 +142,4 @@ int bch2_dev_allocator_start(struct bch_dev *);
int bch2_alloc_write_all(struct bch_fs *, unsigned);
void bch2_fs_allocator_background_init(struct bch_fs *);
void bch2_open_buckets_to_text(struct printbuf *, struct bch_fs *);
#endif /* _BCACHEFS_ALLOC_BACKGROUND_H */

100
fs/bcachefs/alloc_foreground.c
View file

@ -45,7 +45,7 @@
void __bch2_open_bucket_put(struct bch_fs *c, struct open_bucket *ob)
{
struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
if (ob->ec) {
bch2_ec_bucket_written(c, ob);
@ -55,9 +55,9 @@ void __bch2_open_bucket_put(struct bch_fs *c, struct open_bucket *ob)
percpu_down_read(&c->mark_lock);
spin_lock(&ob->lock);
bch2_mark_alloc_bucket(c, ca, PTR_BUCKET_NR(ca, &ob->ptr), false);
bch2_mark_alloc_bucket(c, ca, ob->bucket, false);
ob->valid = false;
ob->type = 0;
ob->data_type = 0;
spin_unlock(&ob->lock);
percpu_up_read(&c->mark_lock);
@ -81,8 +81,7 @@ void bch2_open_bucket_write_error(struct bch_fs *c,
unsigned i;
open_bucket_for_each(c, obs, ob, i)
if (ob->ptr.dev == dev &&
ob->ec)
if (ob->dev == dev && ob->ec)
bch2_ec_bucket_cancel(c, ob);
}
@ -95,18 +94,19 @@ static struct open_bucket *bch2_open_bucket_alloc(struct bch_fs *c)
ob = c->open_buckets + c->open_buckets_freelist;
c->open_buckets_freelist = ob->freelist;
atomic_set(&ob->pin, 1);
ob->type = 0;
ob->data_type = 0;
c->open_buckets_nr_free--;
return ob;
}
static void open_bucket_free_unused(struct bch_fs *c,
struct write_point *wp,
struct open_bucket *ob)
{
struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
bool may_realloc = wp->type == BCH_DATA_user;
struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
bool may_realloc = wp->data_type == BCH_DATA_user;
BUG_ON(ca->open_buckets_partial_nr >
ARRAY_SIZE(ca->open_buckets_partial));
@ -133,11 +133,13 @@ static void verify_not_stale(struct bch_fs *c, const struct open_buckets *obs)
struct open_bucket *ob;
unsigned i;
rcu_read_lock();
open_bucket_for_each(c, obs, ob, i) {
struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
BUG_ON(ptr_stale(ca, &ob->ptr));
BUG_ON(bucket(ca, ob->bucket)->mark.gen != ob->gen);
}
rcu_read_unlock();
#endif
}
@ -246,13 +248,9 @@ struct open_bucket *bch2_bucket_alloc(struct bch_fs *c, struct bch_dev *ca,
ob->valid = true;
ob->sectors_free = ca->mi.bucket_size;
ob->alloc_reserve = reserve;
ob->ptr = (struct bch_extent_ptr) {
.type = 1 << BCH_EXTENT_ENTRY_ptr,
.gen = bucket(ca, b)->mark.gen,
.offset = bucket_to_sector(ca, b),
.dev = ca->dev_idx,
};
ob->dev = ca->dev_idx;
ob->gen = bucket(ca, b)->mark.gen;
ob->bucket = b;
spin_unlock(&ob->lock);
if (c->blocked_allocate_open_bucket) {
@ -333,9 +331,9 @@ static void add_new_bucket(struct bch_fs *c,
struct open_bucket *ob)
{
unsigned durability =
bch_dev_bkey_exists(c, ob->ptr.dev)->mi.durability;
bch_dev_bkey_exists(c, ob->dev)->mi.durability;
__clear_bit(ob->ptr.dev, devs_may_alloc->d);
__clear_bit(ob->dev, devs_may_alloc->d);
*nr_effective += (flags & BUCKET_ALLOC_USE_DURABILITY)
? durability : 1;
*have_cache |= !durability;
@ -445,13 +443,13 @@ static int bucket_alloc_from_stripe(struct bch_fs *c,
continue;
ob = c->open_buckets + h->s->blocks[ec_idx];
if (ob->ptr.dev == devs_sorted.devs[i] &&
if (ob->dev == devs_sorted.devs[i] &&
!test_and_set_bit(ec_idx, h->s->blocks_allocated))
goto got_bucket;
}
goto out_put_head;
got_bucket:
ca = bch_dev_bkey_exists(c, ob->ptr.dev);
ca = bch_dev_bkey_exists(c, ob->dev);
ob->ec_idx = ec_idx;
ob->ec = h->s;
@ -481,12 +479,12 @@ static void get_buckets_from_writepoint(struct bch_fs *c,
unsigned i;
open_bucket_for_each(c, &wp->ptrs, ob, i) {
struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
if (*nr_effective < nr_replicas &&
test_bit(ob->ptr.dev, devs_may_alloc->d) &&
test_bit(ob->dev, devs_may_alloc->d) &&
(ca->mi.durability ||
(wp->type == BCH_DATA_user && !*have_cache)) &&
(wp->data_type == BCH_DATA_user && !*have_cache)) &&
(ob->ec || !need_ec)) {
add_new_bucket(c, ptrs, devs_may_alloc,
nr_effective, have_cache,
@ -518,7 +516,7 @@ static int open_bucket_add_buckets(struct bch_fs *c,
unsigned i;
rcu_read_lock();
devs = target_rw_devs(c, wp->type, target);
devs = target_rw_devs(c, wp->data_type, target);
rcu_read_unlock();
/* Don't allocate from devices we already have pointers to: */
@ -526,7 +524,7 @@ static int open_bucket_add_buckets(struct bch_fs *c,
__clear_bit(devs_have->devs[i], devs.d);
open_bucket_for_each(c, ptrs, ob, i)
__clear_bit(ob->ptr.dev, devs.d);
__clear_bit(ob->dev, devs.d);
if (erasure_code) {
if (!ec_open_bucket(c, ptrs)) {
@ -586,7 +584,7 @@ void bch2_open_buckets_stop_dev(struct bch_fs *c, struct bch_dev *ca,
unsigned i, j;
open_bucket_for_each(c, obs, ob, i) {
bool drop = !ca || ob->ptr.dev == ca->dev_idx;
bool drop = !ca || ob->dev == ca->dev_idx;
if (!drop && ob->ec) {
mutex_lock(&ob->ec->lock);
@ -595,7 +593,7 @@ void bch2_open_buckets_stop_dev(struct bch_fs *c, struct bch_dev *ca,
continue;
ob2 = c->open_buckets + ob->ec->blocks[j];
drop |= ob2->ptr.dev == ca->dev_idx;
drop |= ob2->dev == ca->dev_idx;
}
mutex_unlock(&ob->ec->lock);
}
@ -779,11 +777,11 @@ struct write_point *bch2_alloc_sectors_start(struct bch_fs *c,
wp = writepoint_find(c, write_point.v);
if (wp->type == BCH_DATA_user)
if (wp->data_type == BCH_DATA_user)
ob_flags |= BUCKET_MAY_ALLOC_PARTIAL;
/* metadata may not allocate on cache devices: */
if (wp->type != BCH_DATA_user)
if (wp->data_type != BCH_DATA_user)
have_cache = true;
if (!target || (flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)) {
@ -861,6 +859,20 @@ struct write_point *bch2_alloc_sectors_start(struct bch_fs *c,
}
}
struct bch_extent_ptr bch2_ob_ptr(struct bch_fs *c, struct open_bucket *ob)
{
struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
return (struct bch_extent_ptr) {
.type = 1 << BCH_EXTENT_ENTRY_ptr,
.gen = ob->gen,
.dev = ob->dev,
.offset = bucket_to_sector(ca, ob->bucket) +
ca->mi.bucket_size -
ob->sectors_free,
};
}
/*
* Append pointers to the space we just allocated to @k, and mark @sectors space
* as allocated out of @ob
@ -877,14 +889,13 @@ void bch2_alloc_sectors_append_ptrs(struct bch_fs *c, struct write_point *wp,
wp->sectors_free -= sectors;
open_bucket_for_each(c, &wp->ptrs, ob, i) {
struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
struct bch_extent_ptr ptr = ob->ptr;
struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
struct bch_extent_ptr ptr = bch2_ob_ptr(c, ob);
ptr.cached = cached ||
(!ca->mi.durability &&
wp->type == BCH_DATA_user);
wp->data_type == BCH_DATA_user);
ptr.offset += ca->mi.bucket_size - ob->sectors_free;
bch2_bkey_append_ptr(k, ptr);
BUG_ON(sectors > ob->sectors_free);
@ -915,7 +926,7 @@ static inline void writepoint_init(struct write_point *wp,
enum bch_data_type type)
{
mutex_init(&wp->lock);
wp->type = type;
wp->data_type = type;
}
void bch2_fs_allocator_foreground_init(struct bch_fs *c)
@ -952,3 +963,22 @@ void bch2_fs_allocator_foreground_init(struct bch_fs *c)
writepoint_hash(c, wp->write_point));
}
}
void bch2_open_buckets_to_text(struct printbuf *out, struct bch_fs *c)
{
struct open_bucket *ob;
for (ob = c->open_buckets;
ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
ob++) {
spin_lock(&ob->lock);
if (ob->valid && !ob->on_partial_list) {
pr_buf(out, "%zu ref %u type %s\n",
ob - c->open_buckets,
atomic_read(&ob->pin),
bch2_data_types[ob->data_type]);
}
spin_unlock(&ob->lock);
}
}

5
fs/bcachefs/alloc_foreground.h
View file

@ -85,7 +85,7 @@ static inline void bch2_open_bucket_get(struct bch_fs *c,
unsigned i;
open_bucket_for_each(c, &wp->ptrs, ob, i) {
ob->type = wp->type;
ob->data_type = wp->data_type;
atomic_inc(&ob->pin);
ob_push(c, ptrs, ob);
}
@ -105,6 +105,7 @@ struct write_point *bch2_alloc_sectors_start(struct bch_fs *,
unsigned,
struct closure *);
struct bch_extent_ptr bch2_ob_ptr(struct bch_fs *, struct open_bucket *);
void bch2_alloc_sectors_append_ptrs(struct bch_fs *, struct write_point *,
struct bkey_i *, unsigned, bool);
void bch2_alloc_sectors_done(struct bch_fs *, struct write_point *);
@ -127,4 +128,6 @@ static inline struct write_point_specifier writepoint_ptr(struct write_point *wp
void bch2_fs_allocator_foreground_init(struct bch_fs *);
void bch2_open_buckets_to_text(struct printbuf *, struct bch_fs *);
#endif /* _BCACHEFS_ALLOC_FOREGROUND_H */

9
fs/bcachefs/alloc_types.h
View file

@ -49,12 +49,15 @@ struct open_bucket {
* the block in the stripe this open_bucket corresponds to:
*/
u8 ec_idx;
u8 type;
enum bch_data_type data_type:3;
unsigned valid:1;
unsigned on_partial_list:1;
int alloc_reserve:3;
unsigned sectors_free;
struct bch_extent_ptr ptr;
u8 dev;
u8 gen;
u64 bucket;
struct ec_stripe_new *ec;
};
@ -74,7 +77,7 @@ struct write_point {
struct mutex lock;
u64 last_used;
unsigned long write_point;
enum bch_data_type type;
enum bch_data_type data_type;
/* calculated based on how many pointers we're actually going to use: */
unsigned sectors_free;

8
fs/bcachefs/ec.c
View file

@ -1063,7 +1063,7 @@ void *bch2_writepoint_ec_buf(struct bch_fs *c, struct write_point *wp)
if (!ob)
return NULL;
ca = bch_dev_bkey_exists(c, ob->ptr.dev);
ca = bch_dev_bkey_exists(c, ob->dev);
offset = ca->mi.bucket_size - ob->sectors_free;
return ob->ec->new_stripe.data[ob->ec_idx] + (offset << 9);
@ -1318,7 +1318,7 @@ static int new_stripe_alloc_buckets(struct bch_fs *c, struct ec_stripe_head *h,
BUG_ON(j >= h->s->nr_data + h->s->nr_parity);
h->s->blocks[j] = buckets.v[i];
h->s->new_stripe.key.v.ptrs[j] = ob->ptr;
h->s->new_stripe.key.v.ptrs[j] = bch2_ob_ptr(c, ob);
__set_bit(j, h->s->blocks_gotten);
}
@ -1346,7 +1346,7 @@ static int new_stripe_alloc_buckets(struct bch_fs *c, struct ec_stripe_head *h,
BUG_ON(j >= h->s->nr_data);
h->s->blocks[j] = buckets.v[i];
h->s->new_stripe.key.v.ptrs[j] = ob->ptr;
h->s->new_stripe.key.v.ptrs[j] = bch2_ob_ptr(c, ob);
__set_bit(j, h->s->blocks_gotten);
}
@ -1535,7 +1535,7 @@ void bch2_ec_stop_dev(struct bch_fs *c, struct bch_dev *ca)
continue;
ob = c->open_buckets + h->s->blocks[i];
if (ob->ptr.dev == ca->dev_idx)
if (ob->dev == ca->dev_idx)
goto found;
}
goto unlock;

2
fs/bcachefs/journal.c
View file

@ -789,7 +789,7 @@ static int __bch2_set_nr_journal_buckets(struct bch_dev *ca, unsigned nr,
goto err;
}
b = sector_to_bucket(ca, ob->ptr.offset);
b = ob->bucket;
}
if (c)

3
fs/bcachefs/sysfs.c
View file

@ -10,6 +10,7 @@
#include "bcachefs.h"
#include "alloc_background.h"
#include "alloc_foreground.h"
#include "sysfs.h"
#include "btree_cache.h"
#include "btree_io.h"
@ -723,7 +724,7 @@ static void dev_alloc_debug_to_text(struct printbuf *out, struct bch_dev *ca)
memset(nr, 0, sizeof(nr));
for (i = 0; i < ARRAY_SIZE(c->open_buckets); i++)
nr[c->open_buckets[i].type]++;
nr[c->open_buckets[i].data_type]++;
pr_buf(out,
"\t\t buckets\t sectors fragmented\n"