-2afdc642c2ab4d629993c7f064765ecf25ee483f
+f4b290345a983c534879e603fa5bf4d7465c9e2e
c->pd_controllers_update_seconds * HZ);
}
+/* Persistent alloc info: */
+
static unsigned bch_alloc_val_u64s(const struct bch_alloc *a)
{
unsigned bytes = offsetof(struct bch_alloc, data);
if (a.k->p.offset >= ca->mi.nbuckets)
return;
- g = ca->buckets + a.k->p.offset;
+ lg_local_lock(&c->usage_lock);
+
+ g = bucket(ca, a.k->p.offset);
bucket_cmpxchg(g, new, ({
new.gen = a.v->gen;
new.gen_valid = 1;
g->prio[READ] = get_alloc_field(&d, 2);
if (a.v->fields & (1 << BCH_ALLOC_FIELD_WRITE_TIME))
g->prio[WRITE] = get_alloc_field(&d, 2);
+
+ lg_local_unlock(&c->usage_lock);
}
int bch2_alloc_read(struct bch_fs *c, struct list_head *journal_replay_list)
bch2_alloc_read_key(c, bkey_i_to_s_c(k));
}
- mutex_lock(&c->bucket_lock);
+ mutex_lock(&c->prio_clock[READ].lock);
for_each_member_device(ca, c, i) {
+ down_read(&ca->bucket_lock);
bch2_recalc_min_prio(c, ca, READ);
+ up_read(&ca->bucket_lock);
+ }
+ mutex_unlock(&c->prio_clock[READ].lock);
+
+ mutex_lock(&c->prio_clock[WRITE].lock);
+ for_each_member_device(ca, c, i) {
+ down_read(&ca->bucket_lock);
bch2_recalc_min_prio(c, ca, WRITE);
+ up_read(&ca->bucket_lock);
}
- mutex_unlock(&c->bucket_lock);
+ mutex_unlock(&c->prio_clock[WRITE].lock);
return 0;
}
static int __bch2_alloc_write_key(struct bch_fs *c, struct bch_dev *ca,
- struct bucket *g, struct btree_iter *iter,
+ size_t b, struct btree_iter *iter,
u64 *journal_seq)
{
struct bucket_mark m;
__BKEY_PADDED(k, DIV_ROUND_UP(sizeof(struct bch_alloc), 8)) alloc_key;
+ struct bucket *g;
struct bkey_i_alloc *a;
u8 *d;
int ret;
- bch2_btree_iter_set_pos(iter, POS(ca->dev_idx, g - ca->buckets));
+ bch2_btree_iter_set_pos(iter, POS(ca->dev_idx, b));
do {
ret = bch2_btree_iter_traverse(iter);
if (ret)
break;
+ lg_local_lock(&c->usage_lock);
+ g = bucket(ca, b);
+
/* read mark under btree node lock: */
m = READ_ONCE(g->mark);
a = bkey_alloc_init(&alloc_key.k);
put_alloc_field(&d, 2, g->prio[READ]);
if (a->v.fields & (1 << BCH_ALLOC_FIELD_WRITE_TIME))
put_alloc_field(&d, 2, g->prio[WRITE]);
+ lg_local_unlock(&c->usage_lock);
- bch2_btree_iter_set_pos(iter, a->k.p);
ret = bch2_btree_insert_at(c, NULL, NULL, journal_seq,
BTREE_INSERT_ATOMIC|
BTREE_INSERT_NOFAIL|
int bch2_alloc_replay_key(struct bch_fs *c, struct bpos pos)
{
struct bch_dev *ca;
- struct bucket *g;
struct btree_iter iter;
int ret;
if (pos.offset >= ca->mi.nbuckets)
return 0;
- g = ca->buckets + pos.offset;
-
bch2_btree_iter_init(&iter, c, BTREE_ID_ALLOC, POS_MIN,
BTREE_ITER_INTENT);
- ret = __bch2_alloc_write_key(c, ca, g, &iter, NULL);
+ ret = __bch2_alloc_write_key(c, ca, pos.offset, &iter, NULL);
bch2_btree_iter_unlock(&iter);
return ret;
}
bch2_btree_iter_init(&iter, c, BTREE_ID_ALLOC, POS_MIN,
BTREE_ITER_INTENT);
- for_each_set_bit(bucket, ca->bucket_dirty, ca->mi.nbuckets) {
- ret = __bch2_alloc_write_key(c, ca, ca->buckets + bucket,
- &iter, journal_seq);
+ down_read(&ca->bucket_lock);
+ for_each_set_bit(bucket, ca->buckets_dirty, ca->mi.nbuckets) {
+ ret = __bch2_alloc_write_key(c, ca, bucket, &iter, journal_seq);
if (ret)
break;
- clear_bit(bucket, ca->bucket_dirty);
+ clear_bit(bucket, ca->buckets_dirty);
}
+ up_read(&ca->bucket_lock);
bch2_btree_iter_unlock(&iter);
return ret;
}
-#define BUCKET_GC_GEN_MAX 96U
-
-/**
- * wait_buckets_available - wait on reclaimable buckets
- *
- * If there aren't enough available buckets to fill up free_inc, wait until
- * there are.
- */
-static int wait_buckets_available(struct bch_fs *c, struct bch_dev *ca)
-{
- unsigned long gc_count = c->gc_count;
- int ret = 0;
-
- while (1) {
- set_current_state(TASK_INTERRUPTIBLE);
- if (kthread_should_stop()) {
- ret = -1;
- break;
- }
-
- if (gc_count != c->gc_count)
- ca->inc_gen_really_needs_gc = 0;
-
- if ((ssize_t) (dev_buckets_available(c, ca) -
- ca->inc_gen_really_needs_gc) >=
- (ssize_t) fifo_free(&ca->free_inc))
- break;
-
- up_read(&c->gc_lock);
- schedule();
- try_to_freeze();
- down_read(&c->gc_lock);
- }
-
- __set_current_state(TASK_RUNNING);
- return ret;
-}
-
-static void verify_not_on_freelist(struct bch_fs *c, struct bch_dev *ca,
- size_t bucket)
-{
- if (expensive_debug_checks(c)) {
- size_t iter;
- long i;
- unsigned j;
-
- for (j = 0; j < RESERVE_NR; j++)
- fifo_for_each_entry(i, &ca->free[j], iter)
- BUG_ON(i == bucket);
- fifo_for_each_entry(i, &ca->free_inc, iter)
- BUG_ON(i == bucket);
- }
-}
-
-/* Bucket heap / gen */
+/* Bucket IO clocks: */
static void bch2_recalc_min_prio(struct bch_fs *c, struct bch_dev *ca, int rw)
{
struct prio_clock *clock = &c->prio_clock[rw];
+ struct bucket_array *buckets = bucket_array(ca);
struct bucket *g;
u16 max_delta = 1;
unsigned i;
- lockdep_assert_held(&c->bucket_lock);
+ lockdep_assert_held(&c->prio_clock[rw].lock);
/* Determine min prio for this particular device */
- for_each_bucket(g, ca)
+ for_each_bucket(g, buckets)
max_delta = max(max_delta, (u16) (clock->hand - g->prio[rw]));
ca->min_prio[rw] = clock->hand - max_delta;
static void bch2_rescale_prios(struct bch_fs *c, int rw)
{
struct prio_clock *clock = &c->prio_clock[rw];
+ struct bucket_array *buckets;
struct bch_dev *ca;
struct bucket *g;
unsigned i;
trace_rescale_prios(c);
for_each_member_device(ca, c, i) {
- for_each_bucket(g, ca)
+ down_read(&ca->bucket_lock);
+ buckets = bucket_array(ca);
+
+ for_each_bucket(g, buckets)
g->prio[rw] = clock->hand -
- (clock->hand - g->prio[rw]) / 2;
+ (clock->hand - g->prio[rw]) / 2;
bch2_recalc_min_prio(c, ca, rw);
+
+ up_read(&ca->bucket_lock);
}
}
static void bch2_inc_clock_hand(struct io_timer *timer)
{
struct prio_clock *clock = container_of(timer,
- struct prio_clock, rescale);
+ struct prio_clock, rescale);
struct bch_fs *c = container_of(clock,
- struct bch_fs, prio_clock[clock->rw]);
+ struct bch_fs, prio_clock[clock->rw]);
u64 capacity;
- mutex_lock(&c->bucket_lock);
+ mutex_lock(&clock->lock);
clock->hand++;
if (clock->hand == (u16) (clock->min_prio - 1))
bch2_rescale_prios(c, clock->rw);
- mutex_unlock(&c->bucket_lock);
+ mutex_unlock(&clock->lock);
capacity = READ_ONCE(c->capacity);
static void bch2_prio_timer_init(struct bch_fs *c, int rw)
{
struct prio_clock *clock = &c->prio_clock[rw];
- struct io_timer *timer = &clock->rescale;
- clock->rw = rw;
- timer->fn = bch2_inc_clock_hand;
- timer->expire = c->capacity >> 10;
+ clock->hand = 1;
+ clock->rw = rw;
+ clock->rescale.fn = bch2_inc_clock_hand;
+ clock->rescale.expire = c->capacity >> 10;
+ mutex_init(&clock->lock);
}
+/* Background allocator thread: */
+
/*
- * Background allocation thread: scans for buckets to be invalidated,
- * invalidates them, rewrites prios/gens (marking them as invalidated on disk),
- * then optionally issues discard commands to the newly free buckets, then puts
- * them on the various freelists.
+ * Scans for buckets to be invalidated, invalidates them, rewrites prios/gens
+ * (marking them as invalidated on disk), then optionally issues discard
+ * commands to the newly free buckets, then puts them on the various freelists.
*/
-static inline bool can_inc_bucket_gen(struct bch_dev *ca, struct bucket *g)
+static void verify_not_on_freelist(struct bch_fs *c, struct bch_dev *ca,
+ size_t bucket)
+{
+ if (expensive_debug_checks(c)) {
+ size_t iter;
+ long i;
+ unsigned j;
+
+ for (j = 0; j < RESERVE_NR; j++)
+ fifo_for_each_entry(i, &ca->free[j], iter)
+ BUG_ON(i == bucket);
+ fifo_for_each_entry(i, &ca->free_inc, iter)
+ BUG_ON(i == bucket);
+ }
+}
+
+#define BUCKET_GC_GEN_MAX 96U
+
+/**
+ * wait_buckets_available - wait on reclaimable buckets
+ *
+ * If there aren't enough available buckets to fill up free_inc, wait until
+ * there are.
+ */
+static int wait_buckets_available(struct bch_fs *c, struct bch_dev *ca)
{
- return bucket_gc_gen(ca, g) < BUCKET_GC_GEN_MAX;
+ unsigned long gc_count = c->gc_count;
+ int ret = 0;
+
+ while (1) {
+ set_current_state(TASK_INTERRUPTIBLE);
+ if (kthread_should_stop()) {
+ ret = -1;
+ break;
+ }
+
+ if (gc_count != c->gc_count)
+ ca->inc_gen_really_needs_gc = 0;
+
+ if ((ssize_t) (dev_buckets_available(c, ca) -
+ ca->inc_gen_really_needs_gc) >=
+ (ssize_t) fifo_free(&ca->free_inc))
+ break;
+
+ up_read(&c->gc_lock);
+ schedule();
+ try_to_freeze();
+ down_read(&c->gc_lock);
+ }
+
+ __set_current_state(TASK_RUNNING);
+ return ret;
}
-static bool bch2_can_invalidate_bucket(struct bch_dev *ca, struct bucket *g,
+static bool bch2_can_invalidate_bucket(struct bch_dev *ca,
+ size_t bucket,
struct bucket_mark mark)
{
+ u8 gc_gen;
+
if (!is_available_bucket(mark))
return false;
- if (bucket_gc_gen(ca, g) >= BUCKET_GC_GEN_MAX / 2)
+ gc_gen = bucket_gc_gen(ca, bucket);
+
+ if (gc_gen >= BUCKET_GC_GEN_MAX / 2)
ca->inc_gen_needs_gc++;
- if (bucket_gc_gen(ca, g) >= BUCKET_GC_GEN_MAX)
+ if (gc_gen >= BUCKET_GC_GEN_MAX)
ca->inc_gen_really_needs_gc++;
- return can_inc_bucket_gen(ca, g);
+ return gc_gen < BUCKET_GC_GEN_MAX;
}
static void bch2_invalidate_one_bucket(struct bch_fs *c, struct bch_dev *ca,
- struct bucket *g)
+ size_t bucket)
{
struct bucket_mark m;
spin_lock(&c->freelist_lock);
- if (!bch2_invalidate_bucket(c, ca, g, &m)) {
+ if (!bch2_invalidate_bucket(c, ca, bucket, &m)) {
spin_unlock(&c->freelist_lock);
return;
}
- verify_not_on_freelist(c, ca, g - ca->buckets);
- BUG_ON(!fifo_push(&ca->free_inc, g - ca->buckets));
+ verify_not_on_freelist(c, ca, bucket);
+ BUG_ON(!fifo_push(&ca->free_inc, bucket));
spin_unlock(&c->freelist_lock);
- g->prio[READ] = c->prio_clock[READ].hand;
- g->prio[WRITE] = c->prio_clock[WRITE].hand;
+ /* gc lock held: */
+ bucket_io_clock_reset(c, ca, bucket, READ);
+ bucket_io_clock_reset(c, ca, bucket, WRITE);
if (m.cached_sectors) {
ca->allocator_invalidating_data = true;
*/
static unsigned long bucket_sort_key(struct bch_fs *c, struct bch_dev *ca,
- struct bucket *g, struct bucket_mark m)
+ size_t b, struct bucket_mark m)
{
/*
* Time since last read, scaled to [0, 8) where larger value indicates
* more recently read data:
*/
unsigned long hotness =
- (g->prio[READ] - ca->min_prio[READ]) * 7 /
+ (bucket(ca, b)->prio[READ] - ca->min_prio[READ]) * 7 /
(c->prio_clock[READ].hand - ca->min_prio[READ]);
/* How much we want to keep the data in this bucket: */
(hotness + 1) * bucket_sectors_used(m);
unsigned long needs_journal_commit =
- bucket_needs_journal_commit(m, c->journal.last_seq_ondisk);
+ bucket_needs_journal_commit(m, c->journal.last_seq_ondisk);
return (data_wantness << 9) |
(needs_journal_commit << 8) |
- bucket_gc_gen(ca, g);
+ bucket_gc_gen(ca, b);
}
static inline int bucket_alloc_cmp(alloc_heap *h,
static void invalidate_buckets_lru(struct bch_fs *c, struct bch_dev *ca)
{
+ struct bucket_array *buckets;
struct alloc_heap_entry e;
- struct bucket *g;
+ size_t b;
ca->alloc_heap.used = 0;
- mutex_lock(&c->bucket_lock);
+ mutex_lock(&c->prio_clock[READ].lock);
+ down_read(&ca->bucket_lock);
+
+ buckets = bucket_array(ca);
+
bch2_recalc_min_prio(c, ca, READ);
- bch2_recalc_min_prio(c, ca, WRITE);
/*
* Find buckets with lowest read priority, by building a maxheap sorted
* by read priority and repeatedly replacing the maximum element until
* all buckets have been visited.
*/
- for_each_bucket(g, ca) {
- struct bucket_mark m = READ_ONCE(g->mark);
+ for (b = ca->mi.first_bucket; b < ca->mi.nbuckets; b++) {
+ struct bucket_mark m = READ_ONCE(buckets->b[b].mark);
- if (!bch2_can_invalidate_bucket(ca, g, m))
+ if (!bch2_can_invalidate_bucket(ca, b, m))
continue;
e = (struct alloc_heap_entry) {
- .bucket = g - ca->buckets,
- .key = bucket_sort_key(c, ca, g, m)
+ .bucket = b,
+ .key = bucket_sort_key(c, ca, b, m)
};
heap_add_or_replace(&ca->alloc_heap, e, -bucket_alloc_cmp);
}
+ up_read(&ca->bucket_lock);
+ mutex_unlock(&c->prio_clock[READ].lock);
+
heap_resort(&ca->alloc_heap, bucket_alloc_cmp);
/*
*/
while (!fifo_full(&ca->free_inc) &&
heap_pop(&ca->alloc_heap, e, bucket_alloc_cmp))
- bch2_invalidate_one_bucket(c, ca, &ca->buckets[e.bucket]);
-
- mutex_unlock(&c->bucket_lock);
+ bch2_invalidate_one_bucket(c, ca, e.bucket);
}
static void invalidate_buckets_fifo(struct bch_fs *c, struct bch_dev *ca)
{
+ struct bucket_array *buckets = bucket_array(ca);
struct bucket_mark m;
- struct bucket *g;
- size_t checked = 0;
+ size_t b, checked;
- while (!fifo_full(&ca->free_inc)) {
+ for (checked = 0;
+ checked < ca->mi.nbuckets && !fifo_full(&ca->free_inc);
+ checked++) {
if (ca->fifo_last_bucket < ca->mi.first_bucket ||
ca->fifo_last_bucket >= ca->mi.nbuckets)
ca->fifo_last_bucket = ca->mi.first_bucket;
- g = ca->buckets + ca->fifo_last_bucket++;
- m = READ_ONCE(g->mark);
+ b = ca->fifo_last_bucket++;
- if (bch2_can_invalidate_bucket(ca, g, m))
- bch2_invalidate_one_bucket(c, ca, g);
+ m = READ_ONCE(buckets->b[b].mark);
- if (++checked >= ca->mi.nbuckets)
- return;
+ if (bch2_can_invalidate_bucket(ca, b, m))
+ bch2_invalidate_one_bucket(c, ca, b);
}
}
static void invalidate_buckets_random(struct bch_fs *c, struct bch_dev *ca)
{
+ struct bucket_array *buckets = bucket_array(ca);
struct bucket_mark m;
- struct bucket *g;
- size_t checked = 0;
+ size_t checked;
- while (!fifo_full(&ca->free_inc)) {
- size_t n = bch2_rand_range(ca->mi.nbuckets -
- ca->mi.first_bucket) +
+ for (checked = 0;
+ checked < ca->mi.nbuckets / 2 && !fifo_full(&ca->free_inc);
+ checked++) {
+ size_t b = bch2_rand_range(ca->mi.nbuckets -
+ ca->mi.first_bucket) +
ca->mi.first_bucket;
- g = ca->buckets + n;
- m = READ_ONCE(g->mark);
-
- if (bch2_can_invalidate_bucket(ca, g, m))
- bch2_invalidate_one_bucket(c, ca, g);
+ m = READ_ONCE(buckets->b[b].mark);
- if (++checked >= ca->mi.nbuckets / 2)
- return;
+ if (bch2_can_invalidate_bucket(ca, b, m))
+ bch2_invalidate_one_bucket(c, ca, b);
}
}
ca->inc_gen_really_needs_gc = 0;
switch (ca->mi.replacement) {
- case CACHE_REPLACEMENT_LRU:
- invalidate_buckets_lru(c, ca);
- break;
- case CACHE_REPLACEMENT_FIFO:
- invalidate_buckets_fifo(c, ca);
- break;
- case CACHE_REPLACEMENT_RANDOM:
- invalidate_buckets_random(c, ca);
- break;
+ case CACHE_REPLACEMENT_LRU:
+ invalidate_buckets_lru(c, ca);
+ break;
+ case CACHE_REPLACEMENT_FIFO:
+ invalidate_buckets_fifo(c, ca);
+ break;
+ case CACHE_REPLACEMENT_RANDOM:
+ invalidate_buckets_random(c, ca);
+ break;
}
}
BTREE_ITER_INTENT);
fifo_for_each_entry(b, &ca->free_inc, i) {
- ret = __bch2_alloc_write_key(c, ca, ca->buckets + b,
- &iter, journal_seq);
+ ret = __bch2_alloc_write_key(c, ca, b, &iter, journal_seq);
if (ret)
break;
struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
spin_lock(&ob->lock);
- bch2_mark_alloc_bucket(c, ca, PTR_BUCKET(ca, &ob->ptr), false,
- gc_pos_alloc(c, ob), 0);
+ bch2_mark_alloc_bucket(c, ca, PTR_BUCKET_NR(ca, &ob->ptr),
+ false, gc_pos_alloc(c, ob), 0);
ob->valid = false;
spin_unlock(&ob->lock);
*/
static long bch2_bucket_alloc_startup(struct bch_fs *c, struct bch_dev *ca)
{
- struct bucket *g;
- long r = -1;
+ struct bucket_array *buckets;
+ ssize_t b;
if (!down_read_trylock(&c->gc_lock))
- return r;
+ return -1;
- if (test_bit(BCH_FS_GC_FAILURE, &c->flags))
- goto out;
+ if (test_bit(BCH_FS_GC_FAILURE, &c->flags)) {
+ up_read(&c->gc_lock);
+ return -1;
+ }
- for_each_bucket(g, ca)
- if (!g->mark.touched_this_mount &&
- is_available_bucket(g->mark) &&
- bch2_mark_alloc_bucket_startup(c, ca, g)) {
- r = g - ca->buckets;
- set_bit(r, ca->bucket_dirty);
- break;
+ spin_unlock(&c->freelist_lock);
+
+ down_read(&ca->bucket_lock);
+ buckets = bucket_array(ca);
+
+ spin_lock(&c->freelist_lock);
+
+ for (b = ca->mi.first_bucket; b < ca->mi.nbuckets; b++)
+ if (is_startup_available_bucket(buckets->b[b].mark) &&
+ bch2_mark_alloc_bucket_startup(c, ca, b)) {
+ set_bit(b, ca->buckets_dirty);
+ goto success;
}
-out:
+ b = -1;
+success:
+ up_read(&ca->bucket_lock);
up_read(&c->gc_lock);
- return r;
+ return b;
}
static inline unsigned open_buckets_reserved(enum alloc_reserve reserve)
bool may_alloc_partial,
struct closure *cl)
{
+ struct bucket_array *buckets;
struct open_bucket *ob;
long bucket;
ob = bch2_open_bucket_alloc(c);
spin_lock(&ob->lock);
+ lg_local_lock(&c->usage_lock);
+ buckets = bucket_array(ca);
+
ob->valid = true;
ob->sectors_free = ca->mi.bucket_size;
ob->ptr = (struct bch_extent_ptr) {
- .gen = ca->buckets[bucket].mark.gen,
+ .gen = buckets->b[bucket].mark.gen,
.offset = bucket_to_sector(ca, bucket),
.dev = ca->dev_idx,
};
+
+ bucket_io_clock_reset(c, ca, bucket, READ);
+ bucket_io_clock_reset(c, ca, bucket, WRITE);
+
+ lg_local_unlock(&c->usage_lock);
spin_unlock(&ob->lock);
spin_unlock(&c->freelist_lock);
bch2_wake_allocator(ca);
- ca->buckets[bucket].prio[READ] = c->prio_clock[READ].hand;
- ca->buckets[bucket].prio[WRITE] = c->prio_clock[WRITE].hand;
-
trace_bucket_alloc(ca, reserve);
return ob - c->open_buckets;
}
#include "alloc_types.h"
struct bkey;
-struct bucket;
struct bch_dev;
struct bch_fs;
struct bch_devs_List;
#include <linux/spinlock.h>
#include "clock_types.h"
+#include "fifo.h"
-/*
- * There's two of these clocks, one for reads and one for writes:
- *
- * All fields protected by bucket_lock
- */
+/* There's two of these clocks, one for reads and one for writes: */
struct prio_clock {
/*
* "now" in (read/write) IO time - incremented whenever we do X amount
int rw;
struct io_timer rescale;
+ struct mutex lock;
};
/* There is one reserve for each type of btree, one for prios and gens
RESERVE_NR = 3,
};
+typedef FIFO(long) alloc_fifo;
+
/* Enough for 16 cache devices, 2 tiers and some left over for pipelining */
#define OPEN_BUCKETS_COUNT 256
#define WRITE_POINT_COUNT 32
/* biosets used in cloned bios for writing multiple replicas */
struct bio_set replica_set;
+ /*
+ * Buckets:
+ * Per-bucket arrays are protected by c->usage_lock, bucket_lock and
+ * gc_lock, for device resize - holding any is sufficient for access:
+ * Or rcu_read_lock(), but only for ptr_stale():
+ */
+ struct bucket_array __rcu *buckets;
+ unsigned long *buckets_dirty;
+ /* most out of date gen in the btree */
+ u8 *oldest_gens;
+ struct rw_semaphore bucket_lock;
+
+ struct bch_dev_usage __percpu *usage_percpu;
+ struct bch_dev_usage usage_cached;
+
+ /* Allocator: */
struct task_struct *alloc_thread;
/*
* gens/prios, they'll be moved to the free list (and possibly discarded
* in the process)
*/
- DECLARE_FIFO(long, free)[RESERVE_NR];
- DECLARE_FIFO(long, free_inc);
+ alloc_fifo free[RESERVE_NR];
+ alloc_fifo free_inc;
spinlock_t freelist_lock;
unsigned nr_invalidated;
bool alloc_thread_started;
size_t fifo_last_bucket;
- /* Allocation stuff: */
-
- /* most out of date gen in the btree */
- u8 *oldest_gens;
- struct bucket *buckets;
- unsigned long *bucket_dirty;
-
/* last calculated minimum prio */
u16 min_prio[2];
- /*
- * Bucket book keeping. The first element is updated by GC, the
- * second contains a saved copy of the stats from the beginning
- * of GC.
- */
- struct bch_dev_usage __percpu *usage_percpu;
- struct bch_dev_usage usage_cached;
-
atomic_long_t saturated_count;
size_t inc_gen_needs_gc;
size_t inc_gen_really_needs_gc;
struct bch_fs_usage usage_cached;
struct lglock usage_lock;
- struct mutex bucket_lock;
-
struct closure_waitlist freelist_wait;
/*
#define _BCACHEFS_IOCTL_H
#include <linux/uuid.h>
+#include <asm/ioctl.h>
#include "bcachefs_format.h"
#define BCH_FORCE_IF_DATA_LOST (1 << 0)
#define BCH_BY_INDEX (1 << 4)
+#define BCH_READ_DEV (1 << 5)
+
/* global control dev: */
#define BCH_IOCTL_ASSEMBLE _IOW(0xbc, 1, struct bch_ioctl_assemble)
#define BCH_IOCTL_DISK_EVACUATE _IOW(0xbc, 9, struct bch_ioctl_disk)
#define BCH_IOCTL_DATA _IOW(0xbc, 10, struct bch_ioctl_data)
#define BCH_IOCTL_USAGE _IOWR(0xbc, 11, struct bch_ioctl_usage)
+#define BCH_IOCTL_READ_SUPER _IOW(0xbc, 12, struct bch_ioctl_read_super)
+#define BCH_IOCTL_DISK_GET_IDX _IOW(0xbc, 13, struct bch_ioctl_disk_get_idx)
+#define BCH_IOCTL_DISK_RESIZE _IOW(0xbc, 13, struct bch_ioctl_disk_resize)
struct bch_ioctl_query_uuid {
uuid_le uuid;
struct bch_ioctl_dev_usage devs[0];
};
+struct bch_ioctl_read_super {
+ __u32 flags;
+ __u32 pad;
+ __u64 dev;
+ __u64 size;
+ __u64 sb;
+};
+
+struct bch_ioctl_disk_get_idx {
+ __u64 dev;
+};
+
+struct bch_ioctl_disk_resize {
+ __u32 flags;
+ __u32 pad;
+ __u64 dev;
+ __u64 nbuckets;
+};
+
#endif /* _BCACHEFS_IOCTL_H */
extent_for_each_ptr(e, ptr) {
struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
+ size_t b = PTR_BUCKET_NR(ca, ptr);
struct bucket *g = PTR_BUCKET(ca, ptr);
if (mustfix_fsck_err_on(!g->mark.gen_valid, c,
ptr->gen)) {
g->_mark.gen = ptr->gen;
g->_mark.gen_valid = 1;
- set_bit(g - ca->buckets, ca->bucket_dirty);
+ set_bit(b, ca->buckets_dirty);
}
if (mustfix_fsck_err_on(gen_cmp(ptr->gen, g->mark.gen) > 0, c,
ptr->gen, g->mark.gen)) {
g->_mark.gen = ptr->gen;
g->_mark.gen_valid = 1;
- set_bit(g - ca->buckets, ca->bucket_dirty);
+ set_bit(b, ca->buckets_dirty);
set_bit(BCH_FS_FIXED_GENS, &c->flags);
}
}
}
-
atomic64_set(&c->key_version,
max_t(u64, k.k->version.lo,
atomic64_read(&c->key_version)));
unsigned sectors =
min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
- bch2_mark_metadata_bucket(c, ca, ca->buckets + b,
- type, sectors,
+ bch2_mark_metadata_bucket(c, ca, b, type, sectors,
gc_phase(GC_PHASE_SB), flags);
b++;
start += sectors;
for (i = 0; i < ca->journal.nr; i++) {
b = ca->journal.buckets[i];
- bch2_mark_metadata_bucket(c, ca, ca->buckets + b,
- BCH_DATA_JOURNAL,
+ bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_JOURNAL,
ca->mi.bucket_size,
gc_phase(GC_PHASE_SB), flags);
}
for_each_member_device(ca, c, ci) {
fifo_for_each_entry(i, &ca->free_inc, iter)
- bch2_mark_alloc_bucket(c, ca, &ca->buckets[i], true,
+ bch2_mark_alloc_bucket(c, ca, i, true,
gc_pos_alloc(c, NULL),
BCH_BUCKET_MARK_MAY_MAKE_UNAVAILABLE|
BCH_BUCKET_MARK_GC_LOCK_HELD);
for (j = 0; j < RESERVE_NR; j++)
fifo_for_each_entry(i, &ca->free[j], iter)
- bch2_mark_alloc_bucket(c, ca, &ca->buckets[i], true,
+ bch2_mark_alloc_bucket(c, ca, i, true,
gc_pos_alloc(c, NULL),
BCH_BUCKET_MARK_MAY_MAKE_UNAVAILABLE|
BCH_BUCKET_MARK_GC_LOCK_HELD);
if (ob->valid) {
gc_pos_set(c, gc_pos_alloc(c, ob));
ca = bch_dev_bkey_exists(c, ob->ptr.dev);
- bch2_mark_alloc_bucket(c, ca, PTR_BUCKET(ca, &ob->ptr), true,
+ bch2_mark_alloc_bucket(c, ca, PTR_BUCKET_NR(ca, &ob->ptr), true,
gc_pos_alloc(c, ob),
BCH_BUCKET_MARK_MAY_MAKE_UNAVAILABLE|
BCH_BUCKET_MARK_GC_LOCK_HELD);
static void bch2_gc_start(struct bch_fs *c)
{
struct bch_dev *ca;
- struct bucket *g;
+ struct bucket_array *buckets;
struct bucket_mark new;
unsigned i;
+ size_t b;
int cpu;
lg_global_lock(&c->usage_lock);
lg_global_unlock(&c->usage_lock);
/* Clear bucket marks: */
- for_each_member_device(ca, c, i)
- for_each_bucket(g, ca) {
- bucket_cmpxchg(g, new, ({
+ for_each_member_device(ca, c, i) {
+ down_read(&ca->bucket_lock);
+ buckets = bucket_array(ca);
+
+ for (b = buckets->first_bucket; b < buckets->nbuckets; b++) {
+ bucket_cmpxchg(buckets->b + b, new, ({
new.owned_by_allocator = 0;
new.data_type = 0;
new.cached_sectors = 0;
new.dirty_sectors = 0;
}));
- ca->oldest_gens[g - ca->buckets] = new.gen;
+ ca->oldest_gens[b] = new.gen;
}
+ up_read(&ca->bucket_lock);
+ }
}
/**
return bch2_btree_iter_unlock(&iter) ?: ret;
}
-int bch2_initial_gc(struct bch_fs *c, struct list_head *journal)
+static int __bch2_initial_gc(struct bch_fs *c, struct list_head *journal)
{
unsigned iter = 0;
enum btree_id id;
ret = bch2_journal_mark(c, journal);
if (ret)
- return ret;
+ return ret;
if (test_bit(BCH_FS_FIXED_GENS, &c->flags)) {
if (iter++ > 2) {
return 0;
}
+
+int bch2_initial_gc(struct bch_fs *c, struct list_head *journal)
+{
+ int ret;
+
+ down_write(&c->gc_lock);
+ ret = __bch2_initial_gc(c, journal);
+ up_write(&c->gc_lock);
+
+ return ret;
+}
#include "btree_gc.h"
#include "buckets.h"
#include "error.h"
+#include "movinggc.h"
#include <linux/preempt.h>
#include <trace/events/bcachefs.h>
{
u16 last_seq_ondisk = c->journal.last_seq_ondisk;
struct bch_dev *ca;
+ struct bucket_array *buckets;
struct bucket *g;
struct bucket_mark m;
unsigned i;
- for_each_member_device(ca, c, i)
- for_each_bucket(g, ca) {
+ for_each_member_device(ca, c, i) {
+ down_read(&ca->bucket_lock);
+ buckets = bucket_array(ca);
+
+ for_each_bucket(g, buckets) {
bucket_cmpxchg(g, m, ({
if (!m.journal_seq_valid ||
bucket_needs_journal_commit(m, last_seq_ondisk))
m.journal_seq_valid = 0;
}));
}
+ up_read(&ca->bucket_lock);
+ }
}
#define bch2_usage_add(_acc, _stats) \
}
static void bch2_dev_usage_update(struct bch_fs *c, struct bch_dev *ca,
- struct bucket *g, struct bucket_mark old,
- struct bucket_mark new)
+ struct bucket_mark old, struct bucket_mark new)
{
struct bch_dev_usage *dev_usage;
- BUG_ON((g - ca->buckets) < ca->mi.first_bucket ||
- (g - ca->buckets) >= ca->mi.nbuckets);
+ lockdep_assert_held(&c->usage_lock);
bch2_fs_inconsistent_on(old.data_type && new.data_type &&
old.data_type != new.data_type, c,
"different types of data in same bucket: %u, %u",
old.data_type, new.data_type);
- preempt_disable();
dev_usage = this_cpu_ptr(ca->usage_percpu);
dev_usage->buckets[bucket_type(old)]--;
dev_usage->sectors[new.data_type] += new.dirty_sectors;
dev_usage->sectors[BCH_DATA_CACHED] +=
(int) new.cached_sectors - (int) old.cached_sectors;
- preempt_enable();
if (!is_available_bucket(old) && is_available_bucket(new))
bch2_wake_allocator(ca);
({ \
struct bucket_mark _old = bucket_cmpxchg(g, new, expr); \
\
- bch2_dev_usage_update(c, ca, g, _old, new); \
+ bch2_dev_usage_update(c, ca, _old, new); \
_old; \
})
bool bch2_invalidate_bucket(struct bch_fs *c, struct bch_dev *ca,
- struct bucket *g, struct bucket_mark *old)
+ size_t b, struct bucket_mark *old)
{
+ struct bucket *g;
struct bucket_mark new;
lg_local_lock(&c->usage_lock);
+ g = bucket(ca, b);
+
*old = bucket_data_cmpxchg(c, ca, g, new, ({
if (!is_available_bucket(new)) {
lg_local_unlock(&c->usage_lock);
lg_local_unlock(&c->usage_lock);
if (!old->owned_by_allocator && old->cached_sectors)
- trace_invalidate(ca, bucket_to_sector(ca, g - ca->buckets),
+ trace_invalidate(ca, bucket_to_sector(ca, b),
old->cached_sectors);
return true;
}
bool bch2_mark_alloc_bucket_startup(struct bch_fs *c, struct bch_dev *ca,
- struct bucket *g)
+ size_t b)
{
+ struct bucket *g;
struct bucket_mark new, old;
lg_local_lock(&c->usage_lock);
+ g = bucket(ca, b);
+
old = bucket_data_cmpxchg(c, ca, g, new, ({
- if (new.touched_this_mount ||
- !is_available_bucket(new)) {
+ if (!is_startup_available_bucket(new)) {
lg_local_unlock(&c->usage_lock);
return false;
}
}
void bch2_mark_alloc_bucket(struct bch_fs *c, struct bch_dev *ca,
- struct bucket *g, bool owned_by_allocator,
+ size_t b, bool owned_by_allocator,
struct gc_pos pos, unsigned flags)
{
+ struct bucket *g;
struct bucket_mark old, new;
lg_local_lock(&c->usage_lock);
+ g = bucket(ca, b);
+
if (!(flags & BCH_BUCKET_MARK_GC_LOCK_HELD) &&
gc_will_visit(c, pos)) {
lg_local_unlock(&c->usage_lock);
} while (0)
void bch2_mark_metadata_bucket(struct bch_fs *c, struct bch_dev *ca,
- struct bucket *g, enum bch_data_type type,
+ size_t b, enum bch_data_type type,
unsigned sectors, struct gc_pos pos,
unsigned flags)
{
+ struct bucket *g;
struct bucket_mark old, new;
BUG_ON(!type);
lg_local_lock(&c->usage_lock);
+ g = bucket(ca, b);
+
if (!(flags & BCH_BUCKET_MARK_GC_LOCK_HELD) &&
gc_will_visit(c, pos)) {
lg_local_unlock(&c->usage_lock);
struct bucket_mark old, new;
unsigned saturated;
struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
- struct bucket *g = ca->buckets + PTR_BUCKET_NR(ca, ptr);
+ struct bucket *g = PTR_BUCKET(ca, ptr);
enum bch_data_type data_type = type == S_META
? BCH_DATA_BTREE : BCH_DATA_USER;
u64 v;
old.counter,
new.counter)) != old.counter);
- bch2_dev_usage_update(c, ca, g, old, new);
+ bch2_dev_usage_update(c, ca, old, new);
BUG_ON(!(flags & BCH_BUCKET_MARK_MAY_MAKE_UNAVAILABLE) &&
bucket_became_unavailable(c, old, new));
return bch2_disk_reservation_add(c, res, sectors, flags);
}
+
+/* Startup/shutdown: */
+
+static void buckets_free_rcu(struct rcu_head *rcu)
+{
+ struct bucket_array *buckets =
+ container_of(rcu, struct bucket_array, rcu);
+
+ kvpfree(buckets,
+ sizeof(struct bucket_array) +
+ buckets->nbuckets * sizeof(struct bucket));
+}
+
+int bch2_dev_buckets_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
+{
+ struct bucket_array *buckets = NULL, *old_buckets;
+ unsigned long *buckets_dirty = NULL;
+ u8 *oldest_gens = NULL;
+ alloc_fifo free[RESERVE_NR];
+ alloc_fifo free_inc;
+ alloc_heap alloc_heap;
+ copygc_heap copygc_heap;
+
+ size_t btree_reserve = DIV_ROUND_UP(BTREE_NODE_RESERVE,
+ ca->mi.bucket_size / c->opts.btree_node_size);
+ /* XXX: these should be tunable */
+ size_t reserve_none = max_t(size_t, 4, ca->mi.nbuckets >> 9);
+ size_t copygc_reserve = max_t(size_t, 16, ca->mi.nbuckets >> 7);
+ size_t free_inc_reserve = copygc_reserve / 2;
+ bool resize = ca->buckets != NULL,
+ start_copygc = ca->copygc_thread != NULL;
+ int ret = -ENOMEM;
+ unsigned i;
+
+ memset(&free, 0, sizeof(free));
+ memset(&free_inc, 0, sizeof(free_inc));
+ memset(&alloc_heap, 0, sizeof(alloc_heap));
+ memset(©gc_heap, 0, sizeof(copygc_heap));
+
+ if (!(buckets = kvpmalloc(sizeof(struct bucket_array) +
+ nbuckets * sizeof(struct bucket),
+ GFP_KERNEL|__GFP_ZERO)) ||
+ !(oldest_gens = kvpmalloc(nbuckets * sizeof(u8),
+ GFP_KERNEL|__GFP_ZERO)) ||
+ !(buckets_dirty = kvpmalloc(BITS_TO_LONGS(nbuckets) *
+ sizeof(unsigned long),
+ GFP_KERNEL|__GFP_ZERO)) ||
+ !init_fifo(&free[RESERVE_BTREE], btree_reserve, GFP_KERNEL) ||
+ !init_fifo(&free[RESERVE_MOVINGGC],
+ copygc_reserve, GFP_KERNEL) ||
+ !init_fifo(&free[RESERVE_NONE], reserve_none, GFP_KERNEL) ||
+ !init_fifo(&free_inc, free_inc_reserve, GFP_KERNEL) ||
+ !init_heap(&alloc_heap, free_inc_reserve, GFP_KERNEL) ||
+ !init_heap(©gc_heap, copygc_reserve, GFP_KERNEL))
+ goto err;
+
+ buckets->first_bucket = ca->mi.first_bucket;
+ buckets->nbuckets = nbuckets;
+
+ bch2_copygc_stop(ca);
+
+ down_write(&c->gc_lock);
+ down_write(&ca->bucket_lock);
+ lg_global_lock(&c->usage_lock);
+
+ old_buckets = bucket_array(ca);
+
+ if (resize) {
+ size_t n = min(buckets->nbuckets, old_buckets->nbuckets);
+
+ memcpy(buckets->b,
+ old_buckets->b,
+ n * sizeof(struct bucket));
+ memcpy(oldest_gens,
+ ca->oldest_gens,
+ n * sizeof(u8));
+ memcpy(buckets_dirty,
+ ca->buckets_dirty,
+ BITS_TO_LONGS(n) * sizeof(unsigned long));
+ }
+
+ rcu_assign_pointer(ca->buckets, buckets);
+ buckets = old_buckets;
+
+ swap(ca->oldest_gens, oldest_gens);
+ swap(ca->buckets_dirty, buckets_dirty);
+
+ lg_global_unlock(&c->usage_lock);
+
+ spin_lock(&c->freelist_lock);
+ for (i = 0; i < RESERVE_NR; i++) {
+ fifo_move(&free[i], &ca->free[i]);
+ swap(ca->free[i], free[i]);
+ }
+ fifo_move(&free_inc, &ca->free_inc);
+ swap(ca->free_inc, free_inc);
+ spin_unlock(&c->freelist_lock);
+
+ /* with gc lock held, alloc_heap can't be in use: */
+ swap(ca->alloc_heap, alloc_heap);
+
+ /* and we shut down copygc: */
+ swap(ca->copygc_heap, copygc_heap);
+
+ nbuckets = ca->mi.nbuckets;
+
+ up_write(&ca->bucket_lock);
+ up_write(&c->gc_lock);
+
+ if (start_copygc &&
+ bch2_copygc_start(c, ca))
+ bch_err(ca, "error restarting copygc thread");
+
+ ret = 0;
+err:
+ free_heap(©gc_heap);
+ free_heap(&alloc_heap);
+ free_fifo(&free_inc);
+ for (i = 0; i < RESERVE_NR; i++)
+ free_fifo(&free[i]);
+ kvpfree(buckets_dirty,
+ BITS_TO_LONGS(nbuckets) * sizeof(unsigned long));
+ kvpfree(oldest_gens,
+ nbuckets * sizeof(u8));
+ if (buckets)
+ call_rcu(&old_buckets->rcu, buckets_free_rcu);
+
+ return ret;
+}
+
+void bch2_dev_buckets_free(struct bch_dev *ca)
+{
+ unsigned i;
+
+ free_heap(&ca->copygc_heap);
+ free_heap(&ca->alloc_heap);
+ free_fifo(&ca->free_inc);
+ for (i = 0; i < RESERVE_NR; i++)
+ free_fifo(&ca->free[i]);
+ kvpfree(ca->buckets_dirty,
+ BITS_TO_LONGS(ca->mi.nbuckets) * sizeof(unsigned long));
+ kvpfree(ca->oldest_gens, ca->mi.nbuckets * sizeof(u8));
+ kvpfree(ca->buckets, sizeof(struct bucket_array) +
+ ca->mi.nbuckets * sizeof(struct bucket));
+
+ free_percpu(ca->usage_percpu);
+}
+
+int bch2_dev_buckets_alloc(struct bch_fs *c, struct bch_dev *ca)
+{
+ if (!(ca->usage_percpu = alloc_percpu(struct bch_dev_usage)))
+ return -ENOMEM;
+
+ return bch2_dev_buckets_resize(c, ca, ca->mi.nbuckets);;
+}
#include "buckets_types.h"
#include "super.h"
-#define for_each_bucket(b, ca) \
- for (b = (ca)->buckets + (ca)->mi.first_bucket; \
- b < (ca)->buckets + (ca)->mi.nbuckets; b++)
+#define for_each_bucket(_b, _buckets) \
+ for (_b = (_buckets)->b + (_buckets)->first_bucket; \
+ _b < (_buckets)->b + (_buckets)->nbuckets; _b++)
#define bucket_cmpxchg(g, new, expr) \
({ \
_old; \
})
+static inline struct bucket_array *bucket_array(struct bch_dev *ca)
+{
+ return rcu_dereference_check(ca->buckets,
+ lockdep_is_held(&ca->fs->usage_lock) ||
+ lockdep_is_held(&ca->fs->gc_lock) ||
+ lockdep_is_held(&ca->bucket_lock));
+}
+
+static inline struct bucket *bucket(struct bch_dev *ca, size_t b)
+{
+ struct bucket_array *buckets = bucket_array(ca);
+
+ BUG_ON(b < buckets->first_bucket || b >= buckets->nbuckets);
+ return buckets->b + b;
+}
+
+static inline void bucket_io_clock_reset(struct bch_fs *c, struct bch_dev *ca,
+ size_t b, int rw)
+{
+ bucket(ca, b)->prio[rw] = c->prio_clock[rw].hand;
+}
+
/*
* bucket_gc_gen() returns the difference between the bucket's current gen and
* the oldest gen of any pointer into that bucket in the btree.
*/
-static inline u8 bucket_gc_gen(struct bch_dev *ca, struct bucket *g)
+static inline u8 bucket_gc_gen(struct bch_dev *ca, size_t b)
{
- unsigned long r = g - ca->buckets;
- return g->mark.gen - ca->oldest_gens[r];
+ return bucket(ca, b)->mark.gen - ca->oldest_gens[b];
}
static inline size_t PTR_BUCKET_NR(const struct bch_dev *ca,
return sector_to_bucket(ca, ptr->offset);
}
-static inline struct bucket *PTR_BUCKET(const struct bch_dev *ca,
+static inline struct bucket *PTR_BUCKET(struct bch_dev *ca,
const struct bch_extent_ptr *ptr)
{
- return ca->buckets + PTR_BUCKET_NR(ca, ptr);
+ return bucket(ca, PTR_BUCKET_NR(ca, ptr));
+}
+
+static inline struct bucket_mark ptr_bucket_mark(struct bch_dev *ca,
+ const struct bch_extent_ptr *ptr)
+{
+ struct bucket_mark m;
+
+ rcu_read_lock();
+ m = READ_ONCE(bucket(ca, PTR_BUCKET_NR(ca, ptr))->mark);
+ rcu_read_unlock();
+
+ return m;
}
static inline int gen_cmp(u8 a, u8 b)
* ptr_stale() - check if a pointer points into a bucket that has been
* invalidated.
*/
-static inline u8 ptr_stale(const struct bch_dev *ca,
+static inline u8 ptr_stale(struct bch_dev *ca,
const struct bch_extent_ptr *ptr)
{
- return gen_after(PTR_BUCKET(ca, ptr)->mark.gen, ptr->gen);
+ return gen_after(ptr_bucket_mark(ca, ptr).gen, ptr->gen);
}
/* bucket gc marks */
!mark.nouse);
}
+static inline bool is_startup_available_bucket(struct bucket_mark mark)
+{
+ return !mark.touched_this_mount && is_available_bucket(mark);
+}
+
static inline bool bucket_needs_journal_commit(struct bucket_mark m,
u16 last_seq_ondisk)
{
void bch2_bucket_seq_cleanup(struct bch_fs *);
bool bch2_invalidate_bucket(struct bch_fs *, struct bch_dev *,
- struct bucket *, struct bucket_mark *);
+ size_t, struct bucket_mark *);
bool bch2_mark_alloc_bucket_startup(struct bch_fs *, struct bch_dev *,
- struct bucket *);
+ size_t);
void bch2_mark_alloc_bucket(struct bch_fs *, struct bch_dev *,
- struct bucket *, bool,
- struct gc_pos, unsigned);
+ size_t, bool, struct gc_pos, unsigned);
void bch2_mark_metadata_bucket(struct bch_fs *, struct bch_dev *,
- struct bucket *, enum bch_data_type,
- unsigned, struct gc_pos, unsigned);
+ size_t, enum bch_data_type, unsigned,
+ struct gc_pos, unsigned);
#define BCH_BUCKET_MARK_NOATOMIC (1 << 0)
#define BCH_BUCKET_MARK_MAY_MAKE_UNAVAILABLE (1 << 1)
struct disk_reservation *,
unsigned, int);
+int bch2_dev_buckets_resize(struct bch_fs *, struct bch_dev *, u64);
+void bch2_dev_buckets_free(struct bch_dev *);
+int bch2_dev_buckets_alloc(struct bch_fs *, struct bch_dev *);
+
#endif /* _BUCKETS_H */
};
};
+struct bucket_array {
+ struct rcu_head rcu;
+ u16 first_bucket;
+ size_t nbuckets;
+ struct bucket b[];
+};
+
struct bch_dev_usage {
u64 buckets[BCH_DATA_NR];
u64 buckets_alloc;
return 0;
}
+static long bch2_ioctl_read_super(struct bch_fs *c,
+ struct bch_ioctl_read_super arg)
+{
+ struct bch_dev *ca = NULL;
+ struct bch_sb *sb;
+ int ret = 0;
+
+ if ((arg.flags & ~(BCH_BY_INDEX|BCH_READ_DEV)) ||
+ arg.pad)
+ return -EINVAL;
+
+ mutex_lock(&c->sb_lock);
+
+ if (arg.flags & BCH_READ_DEV) {
+ ca = bch2_device_lookup(c, arg.dev, arg.flags);
+
+ if (IS_ERR(ca)) {
+ ret = PTR_ERR(ca);
+ goto err;
+ }
+
+ sb = ca->disk_sb.sb;
+ } else {
+ sb = c->disk_sb;
+ }
+
+ if (vstruct_bytes(sb) > arg.size) {
+ ret = -ERANGE;
+ goto err;
+ }
+
+ ret = copy_to_user((void __user *)(unsigned long)arg.sb,
+ sb, vstruct_bytes(sb));
+err:
+ if (ca)
+ percpu_ref_put(&ca->ref);
+ mutex_unlock(&c->sb_lock);
+ return ret;
+}
+
+static long bch2_ioctl_disk_get_idx(struct bch_fs *c,
+ struct bch_ioctl_disk_get_idx arg)
+{
+ dev_t dev = huge_decode_dev(arg.dev);
+ struct bch_dev *ca;
+ unsigned i;
+
+ for_each_online_member(ca, c, i)
+ if (ca->disk_sb.bdev->bd_dev == dev) {
+ percpu_ref_put(&ca->io_ref);
+ return i;
+ }
+
+ return -ENOENT;
+}
+
+static long bch2_ioctl_disk_resize(struct bch_fs *c,
+ struct bch_ioctl_disk_resize arg)
+{
+ struct bch_dev *ca;
+ int ret;
+
+ if ((arg.flags & ~BCH_BY_INDEX) ||
+ arg.pad)
+ return -EINVAL;
+
+ ca = bch2_device_lookup(c, arg.dev, arg.flags);
+ if (IS_ERR(ca))
+ return PTR_ERR(ca);
+
+ ret = bch2_dev_resize(c, ca, arg.nbuckets);
+
+ percpu_ref_put(&ca->ref);
+ return ret;
+}
+
#define BCH_IOCTL(_name, _argtype) \
do { \
_argtype i; \
BCH_IOCTL(disk_set_state, struct bch_ioctl_disk_set_state);
case BCH_IOCTL_DISK_EVACUATE:
BCH_IOCTL(disk_evacuate, struct bch_ioctl_disk);
+ case BCH_IOCTL_READ_SUPER:
+ BCH_IOCTL(read_super, struct bch_ioctl_read_super);
+ case BCH_IOCTL_DISK_GET_IDX:
+ BCH_IOCTL(disk_get_idx, struct bch_ioctl_disk_get_idx);
+ case BCH_IOCTL_DISK_RESIZE:
+ BCH_IOCTL(disk_resize, struct bch_ioctl_disk_resize);
default:
return -ENOTTY;
return NULL;
}
+bool bch2_extent_drop_device(struct bkey_s_extent e, unsigned dev)
+{
+ struct bch_extent_ptr *ptr;
+ bool dropped = false;
+
+ extent_for_each_ptr_backwards(e, ptr)
+ if (ptr->dev == dev) {
+ __bch2_extent_drop_ptr(e, ptr);
+ dropped = true;
+ }
+
+ if (dropped)
+ bch2_extent_drop_redundant_crcs(e);
+ return dropped;
+}
+
unsigned bch2_extent_nr_ptrs(struct bkey_s_c_extent e)
{
const struct bch_extent_ptr *ptr;
bch2_extent_drop_redundant_crcs(e);
}
-void bch2_extent_drop_ptr_idx(struct bkey_s_extent e, unsigned idx)
-{
- struct bch_extent_ptr *ptr;
- unsigned i = 0;
-
- extent_for_each_ptr(e, ptr)
- if (i++ == idx)
- goto found;
-
- BUG();
-found:
- bch2_extent_drop_ptr(e, ptr);
-}
-
static inline bool can_narrow_crc(struct bch_extent_crc_unpacked u,
struct bch_extent_crc_unpacked n)
{
unsigned seq;
const char *err;
char buf[160];
- struct bucket *g;
+ struct bucket_mark mark;
struct bch_dev *ca;
unsigned replicas = 0;
bool bad;
extent_for_each_ptr(e, ptr) {
ca = bch_dev_bkey_exists(c, ptr->dev);
- g = PTR_BUCKET(ca, ptr);
replicas++;
if (!test_bit(BCH_FS_ALLOC_READ_DONE, &c->flags))
do {
seq = read_seqcount_begin(&c->gc_pos_lock);
+ mark = ptr_bucket_mark(ca, ptr);
+
bad = gc_pos_cmp(c->gc_pos, gc_pos_btree_node(b)) > 0 &&
- (g->mark.data_type != BCH_DATA_BTREE ||
- g->mark.dirty_sectors < c->opts.btree_node_size);
+ (mark.data_type != BCH_DATA_BTREE ||
+ mark.dirty_sectors < c->opts.btree_node_size);
} while (read_seqcount_retry(&c->gc_pos_lock, seq));
err = "inconsistent";
err:
bch2_bkey_val_to_text(c, btree_node_type(b), buf, sizeof(buf), k);
bch2_fs_bug(c, "%s btree pointer %s: bucket %zi "
- "gen %i last_gc %i mark %08x",
+ "gen %i mark %08x",
err, buf, PTR_BUCKET_NR(ca, ptr),
- PTR_BUCKET(ca, ptr)->mark.gen,
- ca->oldest_gens[PTR_BUCKET_NR(ca, ptr)],
- (unsigned) g->mark.counter);
+ mark.gen, (unsigned) mark.counter);
}
static void bch2_btree_ptr_to_text(struct bch_fs *c, char *buf,
{
const struct bch_extent_ptr *ptr;
struct bch_dev *ca;
- struct bucket *g;
struct bucket_mark mark;
unsigned seq, stale;
char buf[160];
extent_for_each_ptr(e, ptr) {
ca = bch_dev_bkey_exists(c, ptr->dev);
- g = PTR_BUCKET(ca, ptr);
replicas++;
ptrs_per_tier[ca->mi.tier]++;
do {
seq = read_seqcount_begin(&c->gc_pos_lock);
- mark = READ_ONCE(g->mark);
+ mark = ptr_bucket_mark(ca, ptr);
/* between mark and bucket gen */
smp_rmb();
bch2_bkey_val_to_text(c, btree_node_type(b), buf,
sizeof(buf), e.s_c);
bch2_fs_bug(c, "extent pointer bad gc mark: %s:\nbucket %zu "
- "gen %i last_gc %i type %u",
- buf, PTR_BUCKET_NR(ca, ptr), mark.gen,
- ca->oldest_gens[PTR_BUCKET_NR(ca, ptr)],
- mark.data_type);
+ "gen %i type %u", buf,
+ PTR_BUCKET_NR(ca, ptr), mark.gen, mark.data_type);
return;
}
const struct bch_extent_ptr *
bch2_extent_has_device(struct bkey_s_c_extent, unsigned);
+bool bch2_extent_drop_device(struct bkey_s_extent, unsigned);
unsigned bch2_extent_nr_ptrs(struct bkey_s_c_extent);
unsigned bch2_extent_nr_dirty_ptrs(struct bkey_s_c);
void __bch2_extent_drop_ptr(struct bkey_s_extent, struct bch_extent_ptr *);
void bch2_extent_drop_ptr(struct bkey_s_extent, struct bch_extent_ptr *);
-void bch2_extent_drop_ptr_idx(struct bkey_s_extent, unsigned);
bool bch2_cut_front(struct bpos, struct bkey_i *);
bool bch2_cut_back(struct bpos, struct bkey *);
#include "util.h"
-#define DECLARE_FIFO(type, name) \
- struct { \
- size_t front, back, size, mask; \
- type *data; \
- } name
+#define FIFO(type) \
+struct { \
+ size_t front, back, size, mask; \
+ type *data; \
+}
+
+#define DECLARE_FIFO(type, name) FIFO(type) name
#define fifo_buf_size(fifo) \
(roundup_pow_of_two((fifo)->size) * sizeof((fifo)->data[0]))
if (test_bit(ptr->dev, op->failed.d))
bch2_extent_drop_ptr(e, ptr);
- ret = bch2_extent_nr_ptrs(e.c)
- ? bch2_check_mark_super(c, e.c, BCH_DATA_USER)
- : -EIO;
- if (ret) {
- keys->top = keys->keys;
- op->error = ret;
- op->flags |= BCH_WRITE_DONE;
+ if (!bch2_extent_nr_ptrs(e.c)) {
+ ret = -EIO;
goto err;
}
+ if (!(op->flags & BCH_WRITE_NOMARK_REPLICAS)) {
+ ret = bch2_check_mark_super(c, e.c, BCH_DATA_USER);
+ if (ret)
+ goto err;
+ }
+
dst = bkey_next(dst);
}
op->error = ret;
}
}
-err:
+out:
bch2_open_bucket_put_refs(c, &op->open_buckets_nr, op->open_buckets);
if (!(op->flags & BCH_WRITE_DONE))
} else {
continue_at_nobarrier(cl, bch2_write_done, NULL);
}
+ return;
+err:
+ keys->top = keys->keys;
+ op->error = ret;
+ op->flags |= BCH_WRITE_DONE;
+ goto out;
}
static void bch2_write_endio(struct bio *bio)
bch2_extent_crc_append(e, crc);
bch2_alloc_sectors_append_ptrs(op->c, wp, e, crc.compressed_size);
- bkey_extent_set_cached(&e->k, (op->flags & BCH_WRITE_CACHED));
bch2_keylist_push(&op->insert_keys);
}
struct bpos pos = bkey_start_pos(e.k);
int ret = 0;
- PTR_BUCKET(pick->ca, &pick->ptr)->prio[READ] = c->prio_clock[READ].hand;
+ lg_local_lock(&c->usage_lock);
+ bucket_io_clock_reset(c, pick->ca,
+ PTR_BUCKET_NR(pick->ca, &pick->ptr), READ);
+ lg_local_unlock(&c->usage_lock);
narrow_crcs = should_narrow_crcs(e, pick, flags);
BCH_WRITE_PAGES_OWNED = (1 << 5),
BCH_WRITE_ONLY_SPECIFIED_DEVS = (1 << 6),
BCH_WRITE_NOPUT_RESERVATION = (1 << 7),
+ BCH_WRITE_NOMARK_REPLICAS = (1 << 8),
/* Internal: */
- BCH_WRITE_JOURNAL_SEQ_PTR = (1 << 8),
- BCH_WRITE_DONE = (1 << 9),
- BCH_WRITE_LOOPED = (1 << 10),
+ BCH_WRITE_JOURNAL_SEQ_PTR = (1 << 9),
+ BCH_WRITE_DONE = (1 << 10),
+ BCH_WRITE_LOOPED = (1 << 11),
};
static inline u64 *op_journal_seq(struct bch_write_op *op)
ja->nr++;
spin_unlock(&j->lock);
- bch2_mark_metadata_bucket(c, ca, &ca->buckets[bucket],
- BCH_DATA_JOURNAL,
+ bch2_mark_metadata_bucket(c, ca, bucket, BCH_DATA_JOURNAL,
ca->mi.bucket_size,
gc_phase(GC_PHASE_SB), 0);
/**
* journal_next_bucket - move on to the next journal bucket if possible
*/
-static int journal_write_alloc(struct journal *j, unsigned sectors)
+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 bkey_s_extent e = bkey_i_to_s_extent(&j->key);
+ struct bkey_s_extent e;
struct bch_extent_ptr *ptr;
struct journal_device *ja;
struct bch_dev *ca;
READ_ONCE(c->opts.metadata_replicas);
spin_lock(&j->lock);
+ e = bkey_i_to_s_extent(&j->key);
/*
* Drop any pointers to devices that have been removed, are no longer
rcu_read_unlock();
j->prev_buf_sectors = 0;
+
+ bkey_copy(&w->key, &j->key);
spin_unlock(&j->lock);
if (replicas < c->opts.metadata_replicas_required)
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_prev_buf(j);
+ struct bkey_s_c_extent e = bkey_i_to_s_c_extent(&w->key);
+
+ if (!bch2_extent_nr_ptrs(e)) {
+ bch_err(c, "unable to write journal to sufficient devices");
+ goto err;
+ }
+ if (bch2_check_mark_super(c, e, BCH_DATA_JOURNAL))
+ goto err;
+out:
__bch2_time_stats_update(j->write_time, j->write_start_time);
spin_lock(&j->lock);
j->last_seq_ondisk = le64_to_cpu(w->data->last_seq);
+ journal_seq_pin(j, le64_to_cpu(w->data->seq))->devs =
+ bch2_extent_devs(bkey_i_to_s_c_extent(&w->key));
+
/*
* Updating last_seq_ondisk may let journal_reclaim_work() discard more
* buckets:
if (test_bit(JOURNAL_NEED_WRITE, &j->flags))
mod_delayed_work(system_freezable_wq, &j->write_work, 0);
spin_unlock(&j->lock);
-}
-
-static void journal_write_error(struct closure *cl)
-{
- struct journal *j = container_of(cl, struct journal, io);
- struct bch_fs *c = container_of(j, struct bch_fs, journal);
- struct bkey_s_extent e = bkey_i_to_s_extent(&j->key);
-
- while (j->replicas_failed) {
- unsigned idx = __fls(j->replicas_failed);
-
- bch2_extent_drop_ptr_idx(e, idx);
- j->replicas_failed ^= 1 << idx;
- }
-
- if (!bch2_extent_nr_ptrs(e.c)) {
- bch_err(c, "unable to write journal to sufficient devices");
- goto err;
- }
-
- if (bch2_check_mark_super(c, e.c, BCH_DATA_JOURNAL))
- goto err;
-
-out:
- journal_write_done(cl);
return;
err:
bch2_fatal_error(c);
if (bch2_dev_io_err_on(bio->bi_status, ca, "journal write") ||
bch2_meta_write_fault("journal")) {
- /* Was this a flush or an actual journal write? */
- if (ca->journal.ptr_idx != U8_MAX) {
- set_bit(ca->journal.ptr_idx, &j->replicas_failed);
- set_closure_fn(&j->io, journal_write_error,
- system_highpri_wq);
- }
+ struct journal_buf *w = journal_prev_buf(j);
+ unsigned long flags;
+
+ spin_lock_irqsave(&j->err_lock, flags);
+ bch2_extent_drop_device(bkey_i_to_s_extent(&w->key), ca->dev_idx);
+ spin_unlock_irqrestore(&j->err_lock, flags);
}
closure_put(&j->io);
struct jset *jset;
struct bio *bio;
struct bch_extent_ptr *ptr;
- unsigned i, sectors, bytes, ptr_idx = 0;
+ unsigned i, sectors, bytes;
journal_buf_realloc(j, w);
jset = w->data;
bytes = vstruct_bytes(w->data);
memset((void *) w->data + bytes, 0, (sectors << 9) - bytes);
- if (journal_write_alloc(j, sectors)) {
+ if (journal_write_alloc(j, w, sectors)) {
bch2_journal_halt(j);
bch_err(c, "Unable to allocate journal write");
bch2_fatal_error(c);
continue_at(cl, journal_write_done, system_highpri_wq);
}
- if (bch2_check_mark_super(c, bkey_i_to_s_c_extent(&j->key),
- BCH_DATA_JOURNAL))
- goto err;
-
- journal_seq_pin(j, le64_to_cpu(jset->seq))->devs =
- bch2_extent_devs(bkey_i_to_s_c_extent(&j->key));
-
/*
* XXX: we really should just disable the entire journal in nochanges
* mode
if (c->opts.nochanges)
goto no_io;
- extent_for_each_ptr(bkey_i_to_s_extent(&j->key), ptr) {
+ 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 */
this_cpu_add(ca->io_done->sectors[WRITE][BCH_DATA_JOURNAL],
sectors);
- ca->journal.ptr_idx = ptr_idx++;
bio = ca->journal.bio;
bio_reset(bio);
bio->bi_iter.bi_sector = ptr->offset;
for_each_rw_member(ca, c, i)
if (journal_flushes_device(ca) &&
- !bch2_extent_has_device(bkey_i_to_s_c_extent(&j->key), i)) {
+ !bch2_extent_has_device(bkey_i_to_s_c_extent(&w->key), i)) {
percpu_ref_get(&ca->io_ref);
- ca->journal.ptr_idx = U8_MAX;
bio = ca->journal.bio;
bio_reset(bio);
bio->bi_bdev = ca->disk_sb.bdev;
}
no_io:
- extent_for_each_ptr(bkey_i_to_s_extent(&j->key), ptr)
+ extent_for_each_ptr(bkey_i_to_s_extent(&w->key), ptr)
ptr->offset += sectors;
continue_at(cl, journal_write_done, system_highpri_wq);
return ret;
}
-ssize_t bch2_journal_print_debug(struct journal *j, char *buf)
-{
- struct bch_fs *c = container_of(j, struct bch_fs, journal);
- union journal_res_state *s = &j->reservations;
- struct bch_dev *ca;
- unsigned iter;
- ssize_t ret = 0;
-
- rcu_read_lock();
- spin_lock(&j->lock);
-
- ret += scnprintf(buf + ret, PAGE_SIZE - ret,
- "active journal entries:\t%zu\n"
- "seq:\t\t\t%llu\n"
- "last_seq:\t\t%llu\n"
- "last_seq_ondisk:\t%llu\n"
- "reservation count:\t%u\n"
- "reservation offset:\t%u\n"
- "current entry u64s:\t%u\n"
- "io in flight:\t\t%i\n"
- "need write:\t\t%i\n"
- "dirty:\t\t\t%i\n"
- "replay done:\t\t%i\n",
- fifo_used(&j->pin),
- (u64) atomic64_read(&j->seq),
- last_seq(j),
- j->last_seq_ondisk,
- journal_state_count(*s, s->idx),
- s->cur_entry_offset,
- j->cur_entry_u64s,
- s->prev_buf_unwritten,
- test_bit(JOURNAL_NEED_WRITE, &j->flags),
- journal_entry_is_open(j),
- test_bit(JOURNAL_REPLAY_DONE, &j->flags));
-
- for_each_member_device_rcu(ca, c, iter,
- &c->rw_devs[BCH_DATA_JOURNAL]) {
- struct journal_device *ja = &ca->journal;
-
- if (!ja->nr)
- continue;
-
- ret += scnprintf(buf + ret, PAGE_SIZE - ret,
- "dev %u:\n"
- "\tnr\t\t%u\n"
- "\tcur_idx\t\t%u (seq %llu)\n"
- "\tlast_idx\t%u (seq %llu)\n",
- iter, ja->nr,
- ja->cur_idx, ja->bucket_seq[ja->cur_idx],
- ja->last_idx, ja->bucket_seq[ja->last_idx]);
- }
-
- spin_unlock(&j->lock);
- rcu_read_unlock();
-
- return ret;
-}
-
-ssize_t bch2_journal_print_pins(struct journal *j, char *buf)
-{
- struct journal_entry_pin_list *pin_list;
- struct journal_entry_pin *pin;
- ssize_t ret = 0;
- unsigned i;
-
- spin_lock_irq(&j->pin_lock);
- fifo_for_each_entry_ptr(pin_list, &j->pin, i) {
- ret += scnprintf(buf + ret, PAGE_SIZE - ret,
- "%llu: count %u\n",
- journal_pin_seq(j, pin_list),
- atomic_read(&pin_list->count));
-
- list_for_each_entry(pin, &pin_list->list, list)
- ret += scnprintf(buf + ret, PAGE_SIZE - ret,
- "\t%p %pf\n",
- pin, pin->flush);
-
- if (!list_empty(&pin_list->flushed))
- ret += scnprintf(buf + ret, PAGE_SIZE - ret,
- "flushed:\n");
-
- list_for_each_entry(pin, &pin_list->flushed, list)
- ret += scnprintf(buf + ret, PAGE_SIZE - ret,
- "\t%p %pf\n",
- pin, pin->flush);
- }
- spin_unlock_irq(&j->pin_lock);
-
- return ret;
-}
+/* startup/shutdown: */
-static bool bch2_journal_writing_to_device(struct bch_dev *ca)
+static bool bch2_journal_writing_to_device(struct journal *j, unsigned dev_idx)
{
- struct journal *j = &ca->fs->journal;
+ union journal_res_state state;
+ struct journal_buf *w;
bool ret;
spin_lock(&j->lock);
- ret = bch2_extent_has_device(bkey_i_to_s_c_extent(&j->key),
- ca->dev_idx);
+ state = READ_ONCE(j->reservations);
+ w = j->buf + !state.idx;
+
+ ret = state.prev_buf_unwritten &&
+ bch2_extent_has_device(bkey_i_to_s_c_extent(&w->key), dev_idx);
spin_unlock(&j->lock);
return ret;
}
-/*
- * This asumes that ca has already been marked read-only so that
- * journal_next_bucket won't pick buckets out of ca any more.
- * Hence, if the journal is not currently pointing to ca, there
- * will be no new writes to journal entries in ca after all the
- * pending ones have been flushed to disk.
- *
- * If the journal is being written to ca, write a new record, and
- * journal_next_bucket will notice that the device is no longer
- * writeable and pick a new set of devices to write to.
- */
-
-int bch2_journal_move(struct bch_dev *ca)
+void bch2_dev_journal_stop(struct journal *j, struct bch_dev *ca)
{
- struct journal_device *ja = &ca->journal;
- struct journal *j = &ca->fs->journal;
- u64 seq_to_flush = 0;
- unsigned i;
- int ret;
-
- if (bch2_journal_writing_to_device(ca)) {
- /*
- * bch_journal_meta will write a record and we'll wait
- * for the write to complete.
- * Actually writing the journal (journal_write_locked)
- * will call journal_next_bucket which notices that the
- * device is no longer writeable, and picks a new one.
- */
- bch2_journal_meta(j);
- BUG_ON(bch2_journal_writing_to_device(ca));
- }
-
- for (i = 0; i < ja->nr; i++)
- seq_to_flush = max(seq_to_flush, ja->bucket_seq[i]);
-
- bch2_journal_flush_pins(j, seq_to_flush);
-
- /*
- * Force a meta-data journal entry to be written so that
- * we have newer journal entries in devices other than ca,
- * and wait for the meta data write to complete.
- */
- bch2_journal_meta(j);
-
- /*
- * Verify that we no longer need any of the journal entries in
- * the device
- */
spin_lock(&j->lock);
- ret = j->last_seq_ondisk > seq_to_flush ? 0 : -EIO;
+ bch2_extent_drop_device(bkey_i_to_s_extent(&j->key), ca->dev_idx);
spin_unlock(&j->lock);
- return ret;
+ wait_event(j->wait, !bch2_journal_writing_to_device(j, ca->dev_idx));
}
void bch2_fs_journal_stop(struct journal *j)
spin_lock_init(&j->lock);
spin_lock_init(&j->pin_lock);
+ spin_lock_init(&j->err_lock);
init_waitqueue_head(&j->wait);
INIT_DELAYED_WORK(&j->write_work, journal_write_work);
INIT_DELAYED_WORK(&j->reclaim_work, journal_reclaim_work);
return 0;
}
+
+/* debug: */
+
+ssize_t bch2_journal_print_debug(struct journal *j, char *buf)
+{
+ struct bch_fs *c = container_of(j, struct bch_fs, journal);
+ union journal_res_state *s = &j->reservations;
+ struct bch_dev *ca;
+ unsigned iter;
+ ssize_t ret = 0;
+
+ rcu_read_lock();
+ spin_lock(&j->lock);
+
+ ret += scnprintf(buf + ret, PAGE_SIZE - ret,
+ "active journal entries:\t%zu\n"
+ "seq:\t\t\t%llu\n"
+ "last_seq:\t\t%llu\n"
+ "last_seq_ondisk:\t%llu\n"
+ "reservation count:\t%u\n"
+ "reservation offset:\t%u\n"
+ "current entry u64s:\t%u\n"
+ "io in flight:\t\t%i\n"
+ "need write:\t\t%i\n"
+ "dirty:\t\t\t%i\n"
+ "replay done:\t\t%i\n",
+ fifo_used(&j->pin),
+ (u64) atomic64_read(&j->seq),
+ last_seq(j),
+ j->last_seq_ondisk,
+ journal_state_count(*s, s->idx),
+ s->cur_entry_offset,
+ j->cur_entry_u64s,
+ s->prev_buf_unwritten,
+ test_bit(JOURNAL_NEED_WRITE, &j->flags),
+ journal_entry_is_open(j),
+ test_bit(JOURNAL_REPLAY_DONE, &j->flags));
+
+ for_each_member_device_rcu(ca, c, iter,
+ &c->rw_devs[BCH_DATA_JOURNAL]) {
+ struct journal_device *ja = &ca->journal;
+
+ if (!ja->nr)
+ continue;
+
+ ret += scnprintf(buf + ret, PAGE_SIZE - ret,
+ "dev %u:\n"
+ "\tnr\t\t%u\n"
+ "\tcur_idx\t\t%u (seq %llu)\n"
+ "\tlast_idx\t%u (seq %llu)\n",
+ iter, ja->nr,
+ ja->cur_idx, ja->bucket_seq[ja->cur_idx],
+ ja->last_idx, ja->bucket_seq[ja->last_idx]);
+ }
+
+ spin_unlock(&j->lock);
+ rcu_read_unlock();
+
+ return ret;
+}
+
+ssize_t bch2_journal_print_pins(struct journal *j, char *buf)
+{
+ struct journal_entry_pin_list *pin_list;
+ struct journal_entry_pin *pin;
+ ssize_t ret = 0;
+ unsigned i;
+
+ spin_lock_irq(&j->pin_lock);
+ fifo_for_each_entry_ptr(pin_list, &j->pin, i) {
+ ret += scnprintf(buf + ret, PAGE_SIZE - ret,
+ "%llu: count %u\n",
+ journal_pin_seq(j, pin_list),
+ atomic_read(&pin_list->count));
+
+ list_for_each_entry(pin, &pin_list->list, list)
+ ret += scnprintf(buf + ret, PAGE_SIZE - ret,
+ "\t%p %pf\n",
+ pin, pin->flush);
+
+ if (!list_empty(&pin_list->flushed))
+ ret += scnprintf(buf + ret, PAGE_SIZE - ret,
+ "flushed:\n");
+
+ list_for_each_entry(pin, &pin_list->flushed, list)
+ ret += scnprintf(buf + ret, PAGE_SIZE - ret,
+ "\t%p %pf\n",
+ pin, pin->flush);
+ }
+ spin_unlock_irq(&j->pin_lock);
+
+ return ret;
+}
: 0;
}
-int bch2_journal_move(struct bch_dev *);
-
+void bch2_dev_journal_stop(struct journal *, struct bch_dev *);
void bch2_fs_journal_stop(struct journal *);
void bch2_dev_journal_exit(struct bch_dev *);
int bch2_dev_journal_init(struct bch_dev *, struct bch_sb *);
struct journal_buf {
struct jset *data;
+ BKEY_PADDED(key);
+
struct closure_waitlist wait;
unsigned size;
struct closure io;
struct delayed_work write_work;
- unsigned long replicas_failed;
/* Sequence number of most recent journal entry (last entry in @pin) */
atomic64_t seq;
BKEY_PADDED(key);
struct write_point wp;
+ spinlock_t err_lock;
struct delayed_work reclaim_work;
unsigned long last_flushed;
/* Bio for journal reads/writes to this device */
struct bio *bio;
- u8 ptr_idx;
/* for bch_journal_read_device */
struct closure read;
return 0;
mutex_lock(&c->replicas_gc_lock);
- bch2_replicas_gc_start(c,
- (1 << BCH_DATA_JOURNAL)|
- (1 << BCH_DATA_BTREE));
-
- /* 1st, Move the btree nodes off the device */
+ bch2_replicas_gc_start(c, 1 << BCH_DATA_BTREE);
for (i = 0; i < BTREE_ID_NR; i++) {
ret = bch2_move_btree_off(c, ca, i);
if (ret)
goto err;
}
-
- /* There are no prios/gens to move -- they are already in the device. */
-
- /* 2nd. Move the journal off the device */
-
- ret = bch2_journal_move(ca);
- if (ret)
- goto err;
-
err:
bch2_replicas_gc_end(c, ret);
mutex_unlock(&c->replicas_gc_lock);
static int drop_dev_ptrs(struct bch_fs *c, struct bkey_s_extent e,
unsigned dev_idx, int flags, bool metadata)
{
- struct bch_extent_ptr *ptr;
unsigned replicas = metadata ? c->opts.metadata_replicas : c->opts.data_replicas;
unsigned lost = metadata ? BCH_FORCE_IF_METADATA_LOST : BCH_FORCE_IF_DATA_LOST;
unsigned degraded = metadata ? BCH_FORCE_IF_METADATA_DEGRADED : BCH_FORCE_IF_DATA_DEGRADED;
unsigned nr_good;
- extent_for_each_ptr_backwards(e, ptr)
- if (ptr->dev == dev_idx)
- bch2_extent_drop_ptr(e, ptr);
+ bch2_extent_drop_device(e, dev_idx);
nr_good = bch2_extent_nr_good_ptrs(c, e.c);
if ((!nr_good && !(flags & lost)) ||
bch2_extent_normalize(c, extent_i_to_s(insert).s);
bch2_extent_mark_replicas_cached(c, extent_i_to_s(insert));
+ ret = bch2_check_mark_super(c, extent_i_to_s_c(insert),
+ BCH_DATA_USER);
+ if (ret)
+ break;
+
ret = bch2_btree_insert_at(c, &op->res,
NULL, op_journal_seq(op),
BTREE_INSERT_ATOMIC|
m->op.flags |= BCH_WRITE_ONLY_SPECIFIED_DEVS|
BCH_WRITE_PAGES_STABLE|
BCH_WRITE_PAGES_OWNED|
- BCH_WRITE_DATA_ENCODED;
+ BCH_WRITE_DATA_ENCODED|
+ BCH_WRITE_NOMARK_REPLICAS;
m->op.wbio.bio.bi_iter.bi_size = m->op.crc.compressed_size << 9;
m->op.nr_replicas = 1;
{
copygc_heap *h = &ca->copygc_heap;
struct copygc_heap_entry e, *i;
- struct bucket *g;
+ struct bucket_array *buckets;
u64 keys_moved, sectors_moved;
u64 sectors_to_move = 0, sectors_not_moved = 0;
u64 buckets_to_move, buckets_not_moved = 0;
+ size_t b;
int ret;
closure_wait_event(&c->freelist_wait, have_copygc_reserve(ca));
* and repeatedly replacing the maximum element until all
* buckets have been visited.
*/
+ h->used = 0;
/*
* We need bucket marks to be up to date - gc can't be recalculating
* them:
*/
down_read(&c->gc_lock);
- h->used = 0;
- for_each_bucket(g, ca) {
- struct bucket_mark m = READ_ONCE(g->mark);
+ down_read(&ca->bucket_lock);
+ buckets = bucket_array(ca);
+
+ for (b = buckets->first_bucket; b < buckets->nbuckets; b++) {
+ struct bucket_mark m = READ_ONCE(buckets->b[b].mark);
struct copygc_heap_entry e;
if (m.owned_by_allocator ||
continue;
e = (struct copygc_heap_entry) {
- .offset = bucket_to_sector(ca, g - ca->buckets),
+ .offset = bucket_to_sector(ca, b),
.mark = m
};
heap_add_or_replace(h, e, -sectors_used_cmp);
}
+ up_read(&ca->bucket_lock);
up_read(&c->gc_lock);
for (i = h->data; i < h->data + h->used; i++)
&keys_moved,
§ors_moved);
+ down_read(&ca->bucket_lock);
+ buckets = bucket_array(ca);
for (i = h->data; i < h->data + h->used; i++) {
- size_t bucket = sector_to_bucket(ca, i->offset);
- struct bucket_mark m = READ_ONCE(ca->buckets[bucket].mark);
+ size_t b = sector_to_bucket(ca, i->offset);
+ struct bucket_mark m = READ_ONCE(buckets->b[b].mark);
if (i->mark.gen == m.gen && bucket_sectors_used(m)) {
sectors_not_moved += bucket_sectors_used(m);
buckets_not_moved++;
}
}
+ up_read(&ca->bucket_lock);
if (sectors_not_moved && !ret)
bch_warn(c, "copygc finished but %llu/%llu sectors, %llu/%llu buckets not moved",
static void bch2_dev_free(struct bch_dev *);
static int bch2_dev_alloc(struct bch_fs *, unsigned);
-static int bch2_dev_sysfs_online(struct bch_dev *);
+static int bch2_dev_sysfs_online(struct bch_fs *, struct bch_dev *);
static void __bch2_dev_read_only(struct bch_fs *, struct bch_dev *);
struct bch_fs *bch2_bdev_to_fs(struct block_device *bdev)
mutex_init(&c->state_lock);
mutex_init(&c->sb_lock);
mutex_init(&c->replicas_gc_lock);
- mutex_init(&c->bucket_lock);
mutex_init(&c->btree_root_lock);
INIT_WORK(&c->read_only_work, bch2_fs_read_only_work);
seqcount_init(&c->gc_pos_lock);
- c->prio_clock[READ].hand = 1;
- c->prio_clock[READ].min_prio = 0;
- c->prio_clock[WRITE].hand = 1;
- c->prio_clock[WRITE].min_prio = 0;
-
init_waitqueue_head(&c->writeback_wait);
c->writeback_pages_max = (256 << 10) / PAGE_SIZE;
err = "error creating sysfs objects";
__for_each_member_device(ca, c, i, NULL)
- if (bch2_dev_sysfs_online(ca))
+ if (bch2_dev_sysfs_online(c, ca))
goto err;
list_add(&c->list, &bch_fs_list);
static void bch2_dev_free(struct bch_dev *ca)
{
- unsigned i;
-
cancel_work_sync(&ca->io_error_work);
if (ca->kobj.state_in_sysfs &&
free_percpu(ca->io_done);
bioset_exit(&ca->replica_set);
- free_percpu(ca->usage_percpu);
- kvpfree(ca->bucket_dirty, BITS_TO_LONGS(ca->mi.nbuckets) * sizeof(unsigned long));
- kvpfree(ca->buckets, ca->mi.nbuckets * sizeof(struct bucket));
- kvpfree(ca->oldest_gens, ca->mi.nbuckets * sizeof(u8));
- free_heap(&ca->copygc_heap);
- free_heap(&ca->alloc_heap);
- free_fifo(&ca->free_inc);
-
- for (i = 0; i < RESERVE_NR; i++)
- free_fifo(&ca->free[i]);
+ bch2_dev_buckets_free(ca);
percpu_ref_exit(&ca->io_ref);
percpu_ref_exit(&ca->ref);
kobject_put(&ca->kobj);
}
-static void __bch2_dev_offline(struct bch_dev *ca)
+static void __bch2_dev_offline(struct bch_fs *c, struct bch_dev *ca)
{
- struct bch_fs *c = ca->fs;
lockdep_assert_held(&c->state_lock);
complete(&ca->io_ref_completion);
}
-static int bch2_dev_sysfs_online(struct bch_dev *ca)
+static int bch2_dev_sysfs_online(struct bch_fs *c, struct bch_dev *ca)
{
- struct bch_fs *c = ca->fs;
int ret;
if (!c->kobj.state_in_sysfs)
static int bch2_dev_alloc(struct bch_fs *c, unsigned dev_idx)
{
struct bch_member *member;
- size_t reserve_none, movinggc_reserve, free_inc_reserve, total_reserve;
- size_t heap_size;
- unsigned i, btree_node_reserve_buckets;
struct bch_dev *ca;
if (bch2_fs_init_fault("dev_alloc"))
ca->dev_idx = dev_idx;
__set_bit(ca->dev_idx, ca->self.d);
+ init_rwsem(&ca->bucket_lock);
+
writepoint_init(&ca->copygc_write_point, BCH_DATA_USER);
spin_lock_init(&ca->freelist_lock);
ca->uuid = member->uuid;
scnprintf(ca->name, sizeof(ca->name), "dev-%u", dev_idx);
- /* XXX: tune these */
- movinggc_reserve = max_t(size_t, 16, ca->mi.nbuckets >> 7);
- reserve_none = max_t(size_t, 4, ca->mi.nbuckets >> 9);
- /*
- * free_inc must be smaller than the copygc reserve: if it was bigger,
- * one copygc iteration might not make enough buckets available to fill
- * up free_inc and allow the allocator to make forward progress
- */
- free_inc_reserve = movinggc_reserve / 2;
- heap_size = movinggc_reserve * 8;
-
- btree_node_reserve_buckets =
- DIV_ROUND_UP(BTREE_NODE_RESERVE,
- ca->mi.bucket_size / c->opts.btree_node_size);
-
if (percpu_ref_init(&ca->ref, bch2_dev_ref_complete,
0, GFP_KERNEL) ||
percpu_ref_init(&ca->io_ref, bch2_dev_io_ref_complete,
PERCPU_REF_INIT_DEAD, GFP_KERNEL) ||
- !init_fifo(&ca->free[RESERVE_BTREE], btree_node_reserve_buckets,
- GFP_KERNEL) ||
- !init_fifo(&ca->free[RESERVE_MOVINGGC],
- movinggc_reserve, GFP_KERNEL) ||
- !init_fifo(&ca->free[RESERVE_NONE], reserve_none, GFP_KERNEL) ||
- !init_fifo(&ca->free_inc, free_inc_reserve, GFP_KERNEL) ||
- !init_heap(&ca->alloc_heap, free_inc_reserve, GFP_KERNEL) ||
- !init_heap(&ca->copygc_heap,heap_size, GFP_KERNEL) ||
- !(ca->oldest_gens = kvpmalloc(ca->mi.nbuckets *
- sizeof(u8),
- GFP_KERNEL|__GFP_ZERO)) ||
- !(ca->buckets = kvpmalloc(ca->mi.nbuckets *
- sizeof(struct bucket),
- GFP_KERNEL|__GFP_ZERO)) ||
- !(ca->bucket_dirty = kvpmalloc(BITS_TO_LONGS(ca->mi.nbuckets) *
- sizeof(unsigned long),
- GFP_KERNEL|__GFP_ZERO)) ||
- !(ca->usage_percpu = alloc_percpu(struct bch_dev_usage)) ||
+ bch2_dev_buckets_alloc(c, ca) ||
bioset_init(&ca->replica_set, 4,
offsetof(struct bch_write_bio, bio), 0) ||
!(ca->io_done = alloc_percpu(*ca->io_done)))
goto err;
- total_reserve = ca->free_inc.size;
- for (i = 0; i < RESERVE_NR; i++)
- total_reserve += ca->free[i].size;
-
ca->fs = c;
rcu_assign_pointer(c->devs[ca->dev_idx], ca);
- if (bch2_dev_sysfs_online(ca))
+ if (bch2_dev_sysfs_online(c, ca))
pr_warn("error creating sysfs objects");
return 0;
bdevname(ca->disk_sb.bdev, c->name);
bdevname(ca->disk_sb.bdev, ca->name);
- if (bch2_dev_sysfs_online(ca))
- pr_warn("error creating sysfs objects");
-
bch2_mark_dev_superblock(c, ca, BCH_BUCKET_MARK_MAY_MAKE_UNAVAILABLE);
if (ca->mi.state == BCH_MEMBER_STATE_RW)
bch2_copygc_stop(ca);
/*
- * This stops new data writes (e.g. to existing open data
- * buckets) and then waits for all existing writes to
- * complete.
+ * The allocator thread itself allocates btree nodes, so stop it first:
*/
bch2_dev_allocator_stop(ca);
bch2_dev_allocator_remove(c, ca);
+ bch2_dev_journal_stop(&c->journal, ca);
}
static const char *__bch2_dev_read_write(struct bch_fs *c, struct bch_dev *ca)
percpu_ref_put(&ca->ref); /* XXX */
- if (ca->mi.state == BCH_MEMBER_STATE_RW) {
- bch_err(ca, "Cannot remove RW device");
- goto err;
- }
-
if (!bch2_dev_state_allowed(c, ca, BCH_MEMBER_STATE_FAILED, flags)) {
bch_err(ca, "Cannot remove without losing data");
goto err;
}
+ __bch2_dev_read_only(c, ca);
+
/*
* XXX: verify that dev_idx is really not in use anymore, anywhere
*
goto err;
}
- __bch2_dev_offline(ca);
+ __bch2_dev_offline(c, ca);
mutex_lock(&c->sb_lock);
rcu_assign_pointer(c->devs[ca->dev_idx], NULL);
mutex_unlock(&c->state_lock);
return 0;
err:
+ if (ca->mi.state == BCH_MEMBER_STATE_RW)
+ __bch2_dev_read_write(c, ca);
mutex_unlock(&c->state_lock);
return ret;
}
return -EINVAL;
}
- __bch2_dev_offline(ca);
+ __bch2_dev_offline(c, ca);
mutex_unlock(&c->state_lock);
return 0;
mutex_lock(&c->state_lock);
- if (ca->mi.state == BCH_MEMBER_STATE_RW) {
+ if (ca->mi.state == BCH_MEMBER_STATE_RW &&
+ bch2_dev_is_online(ca)) {
bch_err(ca, "Cannot migrate data off RW device");
ret = -EINVAL;
goto err;
return ret;
}
+int bch2_dev_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
+{
+ struct bch_member *mi;
+ int ret = 0;
+
+ mutex_lock(&c->state_lock);
+
+ if (nbuckets < ca->mi.nbuckets) {
+ bch_err(ca, "Cannot shrink yet");
+ ret = -EINVAL;
+ goto err;
+ }
+
+ if (bch2_dev_is_online(ca) &&
+ get_capacity(ca->disk_sb.bdev->bd_disk) <
+ ca->mi.bucket_size * nbuckets) {
+ bch_err(ca, "New size larger than device");
+ ret = -EINVAL;
+ goto err;
+ }
+
+ ret = bch2_dev_buckets_resize(c, ca, nbuckets);
+ if (ret) {
+ bch_err(ca, "Resize error: %i", ret);
+ goto err;
+ }
+
+ mutex_lock(&c->sb_lock);
+ mi = &bch2_sb_get_members(c->disk_sb)->members[ca->dev_idx];
+ mi->nbuckets = cpu_to_le64(nbuckets);
+
+ bch2_write_super(c);
+ mutex_unlock(&c->sb_lock);
+
+ bch2_recalc_capacity(c);
+err:
+ mutex_unlock(&c->state_lock);
+ return ret;
+}
+
/* Filesystem open: */
const char *bch2_fs_open(char * const *devices, unsigned nr_devices,
int bch2_dev_online(struct bch_fs *, const char *);
int bch2_dev_offline(struct bch_fs *, struct bch_dev *, int);
int bch2_dev_evacuate(struct bch_fs *, struct bch_dev *);
+int bch2_dev_resize(struct bch_fs *, struct bch_dev *, u64);
bool bch2_fs_emergency_read_only(struct bch_fs *);
void bch2_fs_read_only(struct bch_fs *);
NULL
};
-typedef unsigned (bucket_map_fn)(struct bch_dev *, struct bucket *, void *);
+typedef unsigned (bucket_map_fn)(struct bch_dev *, size_t, void *);
-static unsigned bucket_priority_fn(struct bch_dev *ca, struct bucket *g,
+static unsigned bucket_priority_fn(struct bch_dev *ca, size_t b,
void *private)
{
+ struct bucket *g = bucket(ca, b);
int rw = (private ? 1 : 0);
return ca->fs->prio_clock[rw].hand - g->prio[rw];
}
-static unsigned bucket_sectors_used_fn(struct bch_dev *ca, struct bucket *g,
+static unsigned bucket_sectors_used_fn(struct bch_dev *ca, size_t b,
void *private)
{
+ struct bucket *g = bucket(ca, b);
return bucket_sectors_used(g->mark);
}
-static unsigned bucket_oldest_gen_fn(struct bch_dev *ca, struct bucket *g,
+static unsigned bucket_oldest_gen_fn(struct bch_dev *ca, size_t b,
void *private)
{
- return bucket_gc_gen(ca, g);
+ return bucket_gc_gen(ca, b);
}
static ssize_t show_quantiles(struct bch_dev *ca, char *buf,
int cmp(const void *l, const void *r)
{ return *((unsigned *) r) - *((unsigned *) l); }
- size_t n = ca->mi.nbuckets, i;
+ size_t i, n;
/* Compute 31 quantiles */
unsigned q[31], *p;
ssize_t ret = 0;
- p = vzalloc(ca->mi.nbuckets * sizeof(unsigned));
- if (!p)
+ down_read(&ca->bucket_lock);
+ n = ca->mi.nbuckets;
+
+ p = vzalloc(n * sizeof(unsigned));
+ if (!p) {
+ up_read(&ca->bucket_lock);
return -ENOMEM;
+ }
for (i = ca->mi.first_bucket; i < n; i++)
- p[i] = fn(ca, &ca->buckets[i], private);
+ p[i] = fn(ca, n, private);
sort(p, n, sizeof(unsigned), cmp, NULL);
+ up_read(&ca->bucket_lock);
while (n &&
!p[n - 1])
projects / bcachefs-tools-debian / commitdiff