-#ifndef _BCACHE_ALLOC_TYPES_H
-#define _BCACHE_ALLOC_TYPES_H
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _BCACHEFS_ALLOC_TYPES_H
+#define _BCACHEFS_ALLOC_TYPES_H
#include <linux/mutex.h>
+#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
- */
-struct prio_clock {
- /*
- * "now" in (read/write) IO time - incremented whenever we do X amount
- * of reads or writes.
- *
- * Goes with the bucket read/write prios: when we read or write to a
- * bucket we reset the bucket's prio to the current hand; thus hand -
- * prio = time since bucket was last read/written.
- *
- * The units are some amount (bytes/sectors) of data read/written, and
- * the units can change on the fly if we need to rescale to fit
- * everything in a u16 - your only guarantee is that the units are
- * consistent.
- */
- u16 hand;
- u16 min_prio;
+struct ec_bucket_buf;
- int rw;
+#define ALLOC_THREAD_STATES() \
+ x(stopped) \
+ x(running) \
+ x(blocked) \
+ x(blocked_full)
- struct io_timer rescale;
+enum allocator_states {
+#define x(n) ALLOCATOR_##n,
+ ALLOC_THREAD_STATES()
+#undef x
};
-/* There is one reserve for each type of btree, one for prios and gens
- * and one for moving GC */
enum alloc_reserve {
- RESERVE_ALLOC = -1,
- RESERVE_BTREE = 0,
- RESERVE_MOVINGGC = 1,
- RESERVE_NONE = 2,
- RESERVE_NR = 3,
+ RESERVE_BTREE_MOVINGGC = -2,
+ RESERVE_BTREE = -1,
+ RESERVE_MOVINGGC = 0,
+ RESERVE_NONE = 1,
+ RESERVE_NR = 2,
};
-struct dev_group {
- spinlock_t lock;
- unsigned nr;
- unsigned cur_device;
- struct {
- u64 weight;
- struct bch_dev *dev;
- } d[BCH_SB_MEMBERS_MAX];
-};
+typedef FIFO(long) alloc_fifo;
+
+#define OPEN_BUCKETS_COUNT 1024
-/* Enough for 16 cache devices, 2 tiers and some left over for pipelining */
-#define OPEN_BUCKETS_COUNT 256
+#define WRITE_POINT_HASH_NR 32
+#define WRITE_POINT_MAX 32
-#define WRITE_POINT_COUNT 16
+typedef u16 open_bucket_idx_t;
struct open_bucket {
- struct list_head list;
- struct mutex lock;
+ spinlock_t lock;
atomic_t pin;
- bool has_full_ptrs;
+ open_bucket_idx_t freelist;
+
/*
- * recalculated every time we allocate from this open_bucket based on
- * how many pointers we're actually going to use:
+ * When an open bucket has an ec_stripe attached, this is the index of
+ * the block in the stripe this open_bucket corresponds to:
*/
+ u8 ec_idx;
+ u8 type;
+ unsigned valid:1;
+ unsigned on_partial_list:1;
+ int alloc_reserve:3;
unsigned sectors_free;
- unsigned nr_ptrs;
- struct bch_extent_ptr ptrs[BCH_REPLICAS_MAX];
- unsigned ptr_offset[BCH_REPLICAS_MAX];
+ struct bch_extent_ptr ptr;
+ struct ec_stripe_new *ec;
+};
+
+#define OPEN_BUCKET_LIST_MAX 15
+
+struct open_buckets {
+ open_bucket_idx_t nr;
+ open_bucket_idx_t v[OPEN_BUCKET_LIST_MAX];
+};
+
+struct dev_stripe_state {
+ u64 next_alloc[BCH_SB_MEMBERS_MAX];
};
struct write_point {
- struct open_bucket *b;
+ struct hlist_node node;
+ struct mutex lock;
+ u64 last_used;
+ unsigned long write_point;
+ enum bch_data_type type;
- /*
- * Throttle writes to this write point if tier 0 is full?
- */
- bool throttle;
+ /* calculated based on how many pointers we're actually going to use: */
+ unsigned sectors_free;
- /*
- * If not NULL, cache group for tiering, promotion and moving GC -
- * always allocates a single replica
- */
- struct dev_group *group;
+ struct open_buckets ptrs;
+ struct dev_stripe_state stripe;
+};
- /*
- * Otherwise do a normal replicated bucket allocation that could come
- * from any device in tier 0 (foreground write)
- */
+struct write_point_specifier {
+ unsigned long v;
};
struct alloc_heap_entry {
size_t bucket;
+ size_t nr;
unsigned long key;
};
typedef HEAP(struct alloc_heap_entry) alloc_heap;
-#endif /* _BCACHE_ALLOC_TYPES_H */
+#endif /* _BCACHEFS_ALLOC_TYPES_H */