bcachefs-tools/libbcachefs/disk_accounting.c

// SPDX-License-Identifier: GPL-2.0

#include "bcachefs.h"
#include "bcachefs_ioctl.h"
#include "btree_cache.h"
#include "btree_journal_iter.h"
#include "btree_update.h"
#include "btree_write_buffer.h"
#include "buckets.h"
#include "compress.h"
#include "disk_accounting.h"
#include "error.h"
#include "journal_io.h"
#include "recovery_passes.h"
#include "replicas.h"

/*
 * Notes on disk accounting:
 *
 * We have two parallel sets of counters to be concerned with, and both must be
 * kept in sync.
 *
 *  - Persistent/on disk accounting, stored in the accounting btree and updated
 *    via btree write buffer updates that treat new accounting keys as deltas to
 *    apply to existing values. But reading from a write buffer btree is
 *    expensive, so we also have
 *
 *  - In memory accounting, where accounting is stored as an array of percpu
 *    counters, indexed by an eytzinger array of disk acounting keys/bpos (which
 *    are the same thing, excepting byte swabbing on big endian).
 *
 *    Cheap to read, but non persistent.
 *
 * Disk accounting updates are generated by transactional triggers; these run as
 * keys enter and leave the btree, and can compare old and new versions of keys;
 * the output of these triggers are deltas to the various counters.
 *
 * Disk accounting updates are done as btree write buffer updates, where the
 * counters in the disk accounting key are deltas that will be applied to the
 * counter in the btree when the key is flushed by the write buffer (or journal
 * replay).
 *
 * To do a disk accounting update:
 * - initialize a disk_accounting_pos, to specify which counter is being update
 * - initialize counter deltas, as an array of 1-3 s64s
 * - call bch2_disk_accounting_mod()
 *
 * This queues up the accounting update to be done at transaction commit time.
 * Underneath, it's a normal btree write buffer update.
 *
 * The transaction commit path is responsible for propagating updates to the in
 * memory counters, with bch2_accounting_mem_mod().
 *
 * The commit path also assigns every disk accounting update a unique version
 * number, based on the journal sequence number and offset within that journal
 * buffer; this is used by journal replay to determine which updates have been
 * done.
 *
 * The transaction commit path also ensures that replicas entry accounting
 * updates are properly marked in the superblock (so that we know whether we can
 * mount without data being unavailable); it will update the superblock if
 * bch2_accounting_mem_mod() tells it to.
 */

static const char * const disk_accounting_type_strs[] = {
#define x(t, n, ...) [n] = #t,
	BCH_DISK_ACCOUNTING_TYPES()
#undef x
	NULL
};

static inline void __accounting_key_init(struct bkey_i *k, struct bpos pos,
					 s64 *d, unsigned nr)
{
	struct bkey_i_accounting *acc = bkey_accounting_init(k);

	acc->k.p = pos;
	set_bkey_val_u64s(&acc->k, sizeof(struct bch_accounting) / sizeof(u64) + nr);

	memcpy_u64s_small(acc->v.d, d, nr);
}

static inline void accounting_key_init(struct bkey_i *k, struct disk_accounting_pos *pos,
				       s64 *d, unsigned nr)
{
	return __accounting_key_init(k, disk_accounting_pos_to_bpos(pos), d, nr);
}

static int bch2_accounting_update_sb_one(struct bch_fs *, struct bpos);

int bch2_disk_accounting_mod(struct btree_trans *trans,
			     struct disk_accounting_pos *k,
			     s64 *d, unsigned nr, bool gc)
{
	BUG_ON(nr > BCH_ACCOUNTING_MAX_COUNTERS);

	/* Normalize: */
	switch (k->type) {
	case BCH_DISK_ACCOUNTING_replicas:
		bubble_sort(k->replicas.devs, k->replicas.nr_devs, u8_cmp);
		break;
	}

	struct bpos pos = disk_accounting_pos_to_bpos(k);

	if (likely(!gc)) {
		struct bkey_i_accounting *a;
#if 0
		for (a = btree_trans_subbuf_base(trans, &trans->accounting);
		     a != btree_trans_subbuf_top(trans, &trans->accounting);
		     a = (void *) bkey_next(&a->k_i))
			if (bpos_eq(a->k.p, pos)) {
				BUG_ON(nr != bch2_accounting_counters(&a->k));
				acc_u64s(a->v.d, d, nr);

				if (bch2_accounting_key_is_zero(accounting_i_to_s_c(a))) {
					unsigned offset = (u64 *) a -
						(u64 *) btree_trans_subbuf_base(trans, &trans->accounting);

					trans->accounting.u64s -= a->k.u64s;
					memmove_u64s_down(a,
							  bkey_next(&a->k_i),
							  trans->accounting.u64s - offset);
				}
				return 0;
			}
#endif
		unsigned u64s = sizeof(*a) / sizeof(u64) + nr;
		a = bch2_trans_subbuf_alloc(trans, &trans->accounting, u64s);
		int ret = PTR_ERR_OR_ZERO(a);
		if (ret)
			return ret;

		__accounting_key_init(&a->k_i, pos, d, nr);
		return 0;
	} else {
		struct { __BKEY_PADDED(k, BCH_ACCOUNTING_MAX_COUNTERS); } k_i;

		__accounting_key_init(&k_i.k, pos, d, nr);

		int ret = bch2_accounting_mem_add(trans, bkey_i_to_s_c_accounting(&k_i.k), true);
		if (ret == -BCH_ERR_btree_insert_need_mark_replicas)
			ret = drop_locks_do(trans,
				bch2_accounting_update_sb_one(trans->c, disk_accounting_pos_to_bpos(k))) ?:
				bch2_accounting_mem_add(trans, bkey_i_to_s_c_accounting(&k_i.k), true);
		return ret;
	}
}

int bch2_mod_dev_cached_sectors(struct btree_trans *trans,
				unsigned dev, s64 sectors,
				bool gc)
{
	struct disk_accounting_pos acc;
	memset(&acc, 0, sizeof(acc));
	acc.type = BCH_DISK_ACCOUNTING_replicas;
	bch2_replicas_entry_cached(&acc.replicas, dev);

	return bch2_disk_accounting_mod(trans, &acc, &sectors, 1, gc);
}

static inline bool is_zero(char *start, char *end)
{
	BUG_ON(start > end);

	for (; start < end; start++)
		if (*start)
			return false;
	return true;
}

#define field_end(p, member)	(((void *) (&p.member)) + sizeof(p.member))

static const unsigned bch2_accounting_type_nr_counters[] = {
#define x(f, id, nr)	[BCH_DISK_ACCOUNTING_##f]	= nr,
	BCH_DISK_ACCOUNTING_TYPES()
#undef x
};

int bch2_accounting_validate(struct bch_fs *c, struct bkey_s_c k,
			     struct bkey_validate_context from)
{
	struct disk_accounting_pos acc_k;
	bpos_to_disk_accounting_pos(&acc_k, k.k->p);
	void *end = &acc_k + 1;
	int ret = 0;

	if (acc_k.type >= BCH_DISK_ACCOUNTING_TYPE_NR)
		return 0;

	bkey_fsck_err_on((from.flags & BCH_VALIDATE_commit) &&
			 bversion_zero(k.k->bversion),
			 c, accounting_key_version_0,
			 "accounting key with version=0");

	switch (acc_k.type) {
	case BCH_DISK_ACCOUNTING_nr_inodes:
		end = field_end(acc_k, nr_inodes);
		break;
	case BCH_DISK_ACCOUNTING_persistent_reserved:
		end = field_end(acc_k, persistent_reserved);
		break;
	case BCH_DISK_ACCOUNTING_replicas:
		bkey_fsck_err_on(!acc_k.replicas.nr_devs,
				 c, accounting_key_replicas_nr_devs_0,
				 "accounting key replicas entry with nr_devs=0");

		bkey_fsck_err_on(acc_k.replicas.nr_required > acc_k.replicas.nr_devs ||
				 (acc_k.replicas.nr_required > 1 &&
				  acc_k.replicas.nr_required == acc_k.replicas.nr_devs),
				 c, accounting_key_replicas_nr_required_bad,
				 "accounting key replicas entry with bad nr_required");

		for (unsigned i = 0; i + 1 < acc_k.replicas.nr_devs; i++)
			bkey_fsck_err_on(acc_k.replicas.devs[i] >= acc_k.replicas.devs[i + 1],
					 c, accounting_key_replicas_devs_unsorted,
					 "accounting key replicas entry with unsorted devs");

		end = (void *) &acc_k.replicas + replicas_entry_bytes(&acc_k.replicas);
		break;
	case BCH_DISK_ACCOUNTING_dev_data_type:
		end = field_end(acc_k, dev_data_type);
		break;
	case BCH_DISK_ACCOUNTING_compression:
		end = field_end(acc_k, compression);
		break;
	case BCH_DISK_ACCOUNTING_snapshot:
		end = field_end(acc_k, snapshot);
		break;
	case BCH_DISK_ACCOUNTING_btree:
		end = field_end(acc_k, btree);
		break;
	case BCH_DISK_ACCOUNTING_rebalance_work:
		end = field_end(acc_k, rebalance_work);
		break;
	}

	bkey_fsck_err_on(!is_zero(end, (void *) (&acc_k + 1)),
			 c, accounting_key_junk_at_end,
			 "junk at end of accounting key");

	const unsigned nr_counters = bch2_accounting_counters(k.k);

	bkey_fsck_err_on(!nr_counters || nr_counters > BCH_ACCOUNTING_MAX_COUNTERS,
			 c, accounting_key_nr_counters_wrong,
			 "accounting key with %u counters, should be %u",
			 nr_counters, bch2_accounting_type_nr_counters[acc_k.type]);
fsck_err:
	return ret;
}

void bch2_accounting_key_to_text(struct printbuf *out, struct disk_accounting_pos *k)
{
	if (k->type >= BCH_DISK_ACCOUNTING_TYPE_NR) {
		prt_printf(out, "unknown type %u", k->type);
		return;
	}

	prt_str(out, disk_accounting_type_strs[k->type]);
	prt_str(out, " ");

	switch (k->type) {
	case BCH_DISK_ACCOUNTING_nr_inodes:
		break;
	case BCH_DISK_ACCOUNTING_persistent_reserved:
		prt_printf(out, "replicas=%u", k->persistent_reserved.nr_replicas);
		break;
	case BCH_DISK_ACCOUNTING_replicas:
		bch2_replicas_entry_to_text(out, &k->replicas);
		break;
	case BCH_DISK_ACCOUNTING_dev_data_type:
		prt_printf(out, "dev=%u data_type=", k->dev_data_type.dev);
		bch2_prt_data_type(out, k->dev_data_type.data_type);
		break;
	case BCH_DISK_ACCOUNTING_compression:
		bch2_prt_compression_type(out, k->compression.type);
		break;
	case BCH_DISK_ACCOUNTING_snapshot:
		prt_printf(out, "id=%u", k->snapshot.id);
		break;
	case BCH_DISK_ACCOUNTING_btree:
		prt_str(out, "btree=");
		bch2_btree_id_to_text(out, k->btree.id);
		break;
	}
}

void bch2_accounting_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
{
	struct bkey_s_c_accounting acc = bkey_s_c_to_accounting(k);
	struct disk_accounting_pos acc_k;
	bpos_to_disk_accounting_pos(&acc_k, k.k->p);

	bch2_accounting_key_to_text(out, &acc_k);

	for (unsigned i = 0; i < bch2_accounting_counters(k.k); i++)
		prt_printf(out, " %lli", acc.v->d[i]);
}

void bch2_accounting_swab(struct bkey_s k)
{
	for (u64 *p = (u64 *) k.v;
	     p < (u64 *) bkey_val_end(k);
	     p++)
		*p = swab64(*p);
}

static inline void __accounting_to_replicas(struct bch_replicas_entry_v1 *r,
					    struct disk_accounting_pos *acc)
{
	unsafe_memcpy(r, &acc->replicas,
		      replicas_entry_bytes(&acc->replicas),
		      "variable length struct");
}

static inline bool accounting_to_replicas(struct bch_replicas_entry_v1 *r, struct bpos p)
{
	struct disk_accounting_pos acc_k;
	bpos_to_disk_accounting_pos(&acc_k, p);

	switch (acc_k.type) {
	case BCH_DISK_ACCOUNTING_replicas:
		__accounting_to_replicas(r, &acc_k);
		return true;
	default:
		return false;
	}
}

static int bch2_accounting_update_sb_one(struct bch_fs *c, struct bpos p)
{
	union bch_replicas_padded r;
	return accounting_to_replicas(&r.e, p)
		? bch2_mark_replicas(c, &r.e)
		: 0;
}

/*
 * Ensure accounting keys being updated are present in the superblock, when
 * applicable (i.e. replicas updates)
 */
int bch2_accounting_update_sb(struct btree_trans *trans)
{
	for (struct bkey_i *i = btree_trans_subbuf_base(trans, &trans->accounting);
	     i != btree_trans_subbuf_top(trans, &trans->accounting);
	     i = bkey_next(i)) {
		int ret = bch2_accounting_update_sb_one(trans->c, i->k.p);
		if (ret)
			return ret;
	}

	return 0;
}

static int __bch2_accounting_mem_insert(struct bch_fs *c, struct bkey_s_c_accounting a)
{
	struct bch_accounting_mem *acc = &c->accounting;

	/* raced with another insert, already present: */
	if (eytzinger0_find(acc->k.data, acc->k.nr, sizeof(acc->k.data[0]),
			    accounting_pos_cmp, &a.k->p) < acc->k.nr)
		return 0;

	struct disk_accounting_pos acc_k;
	bpos_to_disk_accounting_pos(&acc_k, a.k->p);

	struct accounting_mem_entry n = {
		.pos		= a.k->p,
		.bversion	= a.k->bversion,
		.nr_counters	= bch2_accounting_type_nr_counters[acc_k.type],
		.v[0]		= __alloc_percpu_gfp(n.nr_counters * sizeof(u64),
						     sizeof(u64), GFP_KERNEL),
	};

	if (!n.v[0])
		goto err;

	if (acc->gc_running) {
		n.v[1] = __alloc_percpu_gfp(n.nr_counters * sizeof(u64),
					    sizeof(u64), GFP_KERNEL);
		if (!n.v[1])
			goto err;
	}

	if (darray_push(&acc->k, n))
		goto err;

	eytzinger0_sort(acc->k.data, acc->k.nr, sizeof(acc->k.data[0]),
			accounting_pos_cmp, NULL);

	if (trace_accounting_mem_insert_enabled()) {
		CLASS(printbuf, buf)();

		bch2_accounting_to_text(&buf, c, a.s_c);
		trace_accounting_mem_insert(c, buf.buf);
	}
	return 0;
err:
	free_percpu(n.v[1]);
	free_percpu(n.v[0]);
	return bch_err_throw(c, ENOMEM_disk_accounting);
}

int bch2_accounting_mem_insert(struct bch_fs *c, struct bkey_s_c_accounting a,
			       enum bch_accounting_mode mode)
{
	union bch_replicas_padded r;

	if (mode != BCH_ACCOUNTING_read &&
	    accounting_to_replicas(&r.e, a.k->p) &&
	    !bch2_replicas_marked_locked(c, &r.e))
		return bch_err_throw(c, btree_insert_need_mark_replicas);

	percpu_up_read(&c->mark_lock);
	int ret;
	scoped_guard(percpu_write, &c->mark_lock)
		ret = __bch2_accounting_mem_insert(c, a);
	percpu_down_read(&c->mark_lock);
	return ret;
}

int bch2_accounting_mem_insert_locked(struct bch_fs *c, struct bkey_s_c_accounting a,
			       enum bch_accounting_mode mode)
{
	union bch_replicas_padded r;

	if (mode != BCH_ACCOUNTING_read &&
	    accounting_to_replicas(&r.e, a.k->p) &&
	    !bch2_replicas_marked_locked(c, &r.e))
		return bch_err_throw(c, btree_insert_need_mark_replicas);

	return __bch2_accounting_mem_insert(c, a);
}

static bool accounting_mem_entry_is_zero(struct accounting_mem_entry *e)
{
	for (unsigned i = 0; i < e->nr_counters; i++)
		if (percpu_u64_get(e->v[0] + i) ||
		    (e->v[1] &&
		     percpu_u64_get(e->v[1] + i)))
			return false;
	return true;
}

void bch2_accounting_mem_gc(struct bch_fs *c)
{
	struct bch_accounting_mem *acc = &c->accounting;

	guard(percpu_write)(&c->mark_lock);
	struct accounting_mem_entry *dst = acc->k.data;

	darray_for_each(acc->k, src) {
		if (accounting_mem_entry_is_zero(src)) {
			free_percpu(src->v[0]);
			free_percpu(src->v[1]);
		} else {
			*dst++ = *src;
		}
	}

	acc->k.nr = dst - acc->k.data;
	eytzinger0_sort(acc->k.data, acc->k.nr, sizeof(acc->k.data[0]),
			accounting_pos_cmp, NULL);
}

/*
 * Read out accounting keys for replicas entries, as an array of
 * bch_replicas_usage entries.
 *
 * Note: this may be deprecated/removed at smoe point in the future and replaced
 * with something more general, it exists to support the ioctl used by the
 * 'bcachefs fs usage' command.
 */
int bch2_fs_replicas_usage_read(struct bch_fs *c, darray_char *usage)
{
	struct bch_accounting_mem *acc = &c->accounting;
	int ret = 0;

	darray_init(usage);

	guard(percpu_read)(&c->mark_lock);
	darray_for_each(acc->k, i) {
		union {
			u8 bytes[struct_size_t(struct bch_replicas_usage, r.devs,
					       BCH_BKEY_PTRS_MAX)];
			struct bch_replicas_usage r;
		} u;
		u.r.r.nr_devs = BCH_BKEY_PTRS_MAX;

		if (!accounting_to_replicas(&u.r.r, i->pos))
			continue;

		u64 sectors;
		bch2_accounting_mem_read_counters(acc, i - acc->k.data, &sectors, 1, false);
		u.r.sectors = sectors;

		ret = darray_make_room(usage, replicas_usage_bytes(&u.r));
		if (ret)
			break;

		memcpy(&darray_top(*usage), &u.r, replicas_usage_bytes(&u.r));
		usage->nr += replicas_usage_bytes(&u.r);
	}

	if (ret)
		darray_exit(usage);
	return ret;
}

int bch2_fs_accounting_read(struct bch_fs *c, darray_char *out_buf, unsigned accounting_types_mask)
{

	struct bch_accounting_mem *acc = &c->accounting;
	int ret = 0;

	darray_init(out_buf);

	guard(percpu_read)(&c->mark_lock);
	darray_for_each(acc->k, i) {
		struct disk_accounting_pos a_p;
		bpos_to_disk_accounting_pos(&a_p, i->pos);

		if (!(accounting_types_mask & BIT(a_p.type)))
			continue;

		ret = darray_make_room(out_buf, sizeof(struct bkey_i_accounting) +
				       sizeof(u64) * i->nr_counters);
		if (ret)
			break;

		struct bkey_i_accounting *a_out =
			bkey_accounting_init((void *) &darray_top(*out_buf));
		set_bkey_val_u64s(&a_out->k, i->nr_counters);
		a_out->k.p = i->pos;
		bch2_accounting_mem_read_counters(acc, i - acc->k.data,
						  a_out->v.d, i->nr_counters, false);

		if (!bch2_accounting_key_is_zero(accounting_i_to_s_c(a_out)))
			out_buf->nr += bkey_bytes(&a_out->k);
	}

	if (ret)
		darray_exit(out_buf);
	return ret;
}

static void bch2_accounting_free_counters(struct bch_accounting_mem *acc, bool gc)
{
	darray_for_each(acc->k, e) {
		free_percpu(e->v[gc]);
		e->v[gc] = NULL;
	}
}

int bch2_gc_accounting_start(struct bch_fs *c)
{
	struct bch_accounting_mem *acc = &c->accounting;
	int ret = 0;

	guard(percpu_write)(&c->mark_lock);
	darray_for_each(acc->k, e) {
		e->v[1] = __alloc_percpu_gfp(e->nr_counters * sizeof(u64),
					     sizeof(u64), GFP_KERNEL);
		if (!e->v[1]) {
			bch2_accounting_free_counters(acc, true);
			ret = bch_err_throw(c, ENOMEM_disk_accounting);
			break;
		}
	}

	acc->gc_running = !ret;
	return ret;
}

int bch2_gc_accounting_done(struct bch_fs *c)
{
	struct bch_accounting_mem *acc = &c->accounting;
	CLASS(btree_trans, trans)(c);
	CLASS(printbuf, buf)();
	struct bpos pos = POS_MIN;
	int ret = 0;

	guard(percpu_write)(&c->mark_lock);
	while (1) {
		unsigned idx = eytzinger0_find_ge(acc->k.data, acc->k.nr, sizeof(acc->k.data[0]),
						  accounting_pos_cmp, &pos);

		if (idx >= acc->k.nr)
			break;

		struct accounting_mem_entry *e = acc->k.data + idx;
		pos = bpos_successor(e->pos);

		struct disk_accounting_pos acc_k;
		bpos_to_disk_accounting_pos(&acc_k, e->pos);

		if (acc_k.type >= BCH_DISK_ACCOUNTING_TYPE_NR)
			continue;

		u64 src_v[BCH_ACCOUNTING_MAX_COUNTERS];
		u64 dst_v[BCH_ACCOUNTING_MAX_COUNTERS];

		unsigned nr = e->nr_counters;
		bch2_accounting_mem_read_counters(acc, idx, dst_v, nr, false);
		bch2_accounting_mem_read_counters(acc, idx, src_v, nr, true);

		if (memcmp(dst_v, src_v, nr * sizeof(u64))) {
			printbuf_reset(&buf);
			prt_str(&buf, "accounting mismatch for ");
			bch2_accounting_key_to_text(&buf, &acc_k);

			prt_str(&buf, ":\n      got");
			for (unsigned j = 0; j < nr; j++)
				prt_printf(&buf, " %llu", dst_v[j]);

			prt_str(&buf,  "\nshould be");
			for (unsigned j = 0; j < nr; j++)
				prt_printf(&buf, " %llu", src_v[j]);

			for (unsigned j = 0; j < nr; j++)
				src_v[j] -= dst_v[j];

			bch2_trans_unlock_long(trans);

			if (fsck_err(c, accounting_mismatch, "%s", buf.buf)) {
				percpu_up_write(&c->mark_lock);
				ret = commit_do(trans, NULL, NULL,
						BCH_TRANS_COMMIT_skip_accounting_apply,
						bch2_disk_accounting_mod(trans, &acc_k, src_v, nr, false));
				percpu_down_write(&c->mark_lock);
				if (ret)
					goto err;

				if (!test_bit(BCH_FS_may_go_rw, &c->flags)) {
					memset(&trans->fs_usage_delta, 0, sizeof(trans->fs_usage_delta));
					struct { __BKEY_PADDED(k, BCH_ACCOUNTING_MAX_COUNTERS); } k_i;

					accounting_key_init(&k_i.k, &acc_k, src_v, nr);
					bch2_accounting_mem_mod_locked(trans,
								bkey_i_to_s_c_accounting(&k_i.k),
								BCH_ACCOUNTING_normal, true);

					guard(preempt)();
					struct bch_fs_usage_base *dst = this_cpu_ptr(c->usage);
					struct bch_fs_usage_base *src = &trans->fs_usage_delta;
					acc_u64s((u64 *) dst, (u64 *) src, sizeof(*src) / sizeof(u64));
				}
			}
		}
	}
err:
fsck_err:
	bch_err_fn(c, ret);
	return ret;
}

static int accounting_read_key(struct btree_trans *trans, struct bkey_s_c k)
{
	struct bch_fs *c = trans->c;

	if (k.k->type != KEY_TYPE_accounting)
		return 0;

	guard(percpu_read)(&c->mark_lock);
	return bch2_accounting_mem_mod_locked(trans, bkey_s_c_to_accounting(k),
					      BCH_ACCOUNTING_read, false);
}

static int bch2_disk_accounting_validate_late(struct btree_trans *trans,
					      struct disk_accounting_pos *acc,
					      u64 *v, unsigned nr)
{
	struct bch_fs *c = trans->c;
	CLASS(printbuf, buf)();
	int ret = 0, invalid_dev = -1;

	switch (acc->type) {
	case BCH_DISK_ACCOUNTING_replicas: {
		union bch_replicas_padded r;
		__accounting_to_replicas(&r.e, acc);

		for (unsigned i = 0; i < r.e.nr_devs; i++)
			if (r.e.devs[i] != BCH_SB_MEMBER_INVALID &&
			    !bch2_dev_exists(c, r.e.devs[i])) {
				invalid_dev = r.e.devs[i];
				goto invalid_device;
			}

		/*
		 * All replicas entry checks except for invalid device are done
		 * in bch2_accounting_validate
		 */
		BUG_ON(bch2_replicas_entry_validate(&r.e, c, &buf));

		if (fsck_err_on(!bch2_replicas_marked_locked(c, &r.e),
				trans, accounting_replicas_not_marked,
				"accounting not marked in superblock replicas\n%s",
				(printbuf_reset(&buf),
				 bch2_accounting_key_to_text(&buf, acc),
				 buf.buf))) {
			/*
			 * We're not RW yet and still single threaded, dropping
			 * and retaking lock is ok:
			 */
			percpu_up_write(&c->mark_lock);
			ret = bch2_mark_replicas(c, &r.e);
			if (ret)
				goto fsck_err;
			percpu_down_write(&c->mark_lock);
		}
		break;
	}

	case BCH_DISK_ACCOUNTING_dev_data_type:
		if (!bch2_dev_exists(c, acc->dev_data_type.dev)) {
			invalid_dev = acc->dev_data_type.dev;
			goto invalid_device;
		}
		break;
	}

fsck_err:
	return ret;
invalid_device:
	if (fsck_err(trans, accounting_to_invalid_device,
		     "accounting entry points to invalid device %i\n%s",
		     invalid_dev,
		     (printbuf_reset(&buf),
		      bch2_accounting_key_to_text(&buf, acc),
		      buf.buf))) {
		for (unsigned i = 0; i < nr; i++)
			v[i] = -v[i];

		ret = commit_do(trans, NULL, NULL, 0,
				bch2_disk_accounting_mod(trans, acc, v, nr, false)) ?:
			-BCH_ERR_remove_disk_accounting_entry;
	} else {
		ret = bch_err_throw(c, remove_disk_accounting_entry);
	}
	goto fsck_err;
}

static struct journal_key *accumulate_newer_accounting_keys(struct btree_trans *trans, struct journal_key *i)
{
	struct bch_fs *c = trans->c;
	struct journal_keys *keys = &c->journal_keys;
	struct bkey_i *k = journal_key_k(c, i);
	int ret = 0;

	darray_for_each_from(*keys, j, i + 1) {
		if (journal_key_cmp(c, i, j))
			return j;

		struct bkey_i *n = journal_key_k(c, j);
		if (n->k.type == KEY_TYPE_accounting) {
			if (bversion_cmp(k->k.bversion, n->k.bversion) >= 0) {
				CLASS(printbuf, buf)();
				prt_printf(&buf, "accounting keys with out of order versions:");

				prt_newline(&buf);
				prt_printf(&buf, "%u.%u ", i->journal_seq_offset, i->journal_offset);
				bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(k));
				prt_newline(&buf);
				prt_printf(&buf, "%u.%u ", j->journal_seq_offset, j->journal_offset);
				bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(n));
				fsck_err(trans, accounting_key_version_out_of_order, "%s", buf.buf);
			}

			bch2_accounting_accumulate(bkey_i_to_accounting(k),
						   bkey_i_to_s_c_accounting(n));
			j->overwritten = true;
		}
	}

	return &darray_top(*keys);
fsck_err:
	return ERR_PTR(ret);
}

static struct journal_key *accumulate_and_read_journal_accounting(struct btree_trans *trans, struct journal_key *i)
{
	struct journal_key *next = accumulate_newer_accounting_keys(trans, i);

	int ret = PTR_ERR_OR_ZERO(next) ?:
		accounting_read_key(trans, bkey_i_to_s_c(journal_key_k(trans->c, i)));
	return ret ? ERR_PTR(ret) : next;
}

static int accounting_read_mem_fixups(struct btree_trans *trans)
{
	struct bch_fs *c = trans->c;
	struct bch_accounting_mem *acc = &c->accounting;
	CLASS(printbuf, underflow_err)();

	scoped_guard(percpu_write, &c->mark_lock) {
		darray_for_each_reverse(acc->k, i) {
			struct disk_accounting_pos acc_k;
			bpos_to_disk_accounting_pos(&acc_k, i->pos);

			u64 v[BCH_ACCOUNTING_MAX_COUNTERS];
			memset(v, 0, sizeof(v));

			for (unsigned j = 0; j < i->nr_counters; j++)
				v[j] = percpu_u64_get(i->v[0] + j);

			/*
			 * If the entry counters are zeroed, it should be treated as
			 * nonexistent - it might point to an invalid device.
			 *
			 * Remove it, so that if it's re-added it gets re-marked in the
			 * superblock:
			 */
			int ret = bch2_is_zero(v, sizeof(v[0]) * i->nr_counters)
				? -BCH_ERR_remove_disk_accounting_entry
				: bch2_disk_accounting_validate_late(trans, &acc_k, v, i->nr_counters);

			if (ret == -BCH_ERR_remove_disk_accounting_entry) {
				free_percpu(i->v[0]);
				free_percpu(i->v[1]);
				darray_remove_item(&acc->k, i);
				ret = 0;
				continue;
			}

			if (ret)
				return ret;
		}

		eytzinger0_sort(acc->k.data, acc->k.nr, sizeof(acc->k.data[0]),
				accounting_pos_cmp, NULL);

		for (unsigned i = 0; i < acc->k.nr; i++) {
			struct disk_accounting_pos k;
			bpos_to_disk_accounting_pos(&k, acc->k.data[i].pos);

			u64 v[BCH_ACCOUNTING_MAX_COUNTERS];
			bch2_accounting_mem_read_counters(acc, i, v, ARRAY_SIZE(v), false);

			/*
			 * Check for underflow, schedule check_allocations
			 * necessary:
			 *
			 * XXX - see if we can factor this out to run on a bkey
			 * so we can check everything lazily, right now we don't
			 * check the non in-mem counters at all
			 */
			bool underflow = false;
			for (unsigned j = 0; j < acc->k.data[i].nr_counters; j++)
				underflow |= (s64) v[j] < 0;

			if (underflow) {
				if (!underflow_err.pos) {
					bch2_log_msg_start(c, &underflow_err);
					prt_printf(&underflow_err, "Accounting underflow for\n");
				}
				bch2_accounting_key_to_text(&underflow_err, &k);

				for (unsigned j = 0; j < acc->k.data[i].nr_counters; j++)
					prt_printf(&underflow_err, " %lli", v[j]);
				prt_newline(&underflow_err);
			}

			guard(preempt)();
			struct bch_fs_usage_base *usage = this_cpu_ptr(c->usage);

			switch (k.type) {
			case BCH_DISK_ACCOUNTING_persistent_reserved:
				usage->reserved += v[0] * k.persistent_reserved.nr_replicas;
				break;
			case BCH_DISK_ACCOUNTING_replicas:
				fs_usage_data_type_to_base(usage, k.replicas.data_type, v[0]);
				break;
			case BCH_DISK_ACCOUNTING_dev_data_type: {
				guard(rcu)();
				struct bch_dev *ca = bch2_dev_rcu_noerror(c, k.dev_data_type.dev);
				if (ca) {
					struct bch_dev_usage_type __percpu *d = &ca->usage->d[k.dev_data_type.data_type];
					percpu_u64_set(&d->buckets,	v[0]);
					percpu_u64_set(&d->sectors,	v[1]);
					percpu_u64_set(&d->fragmented,	v[2]);

					if (k.dev_data_type.data_type == BCH_DATA_sb ||
					    k.dev_data_type.data_type == BCH_DATA_journal)
						usage->hidden += v[0] * ca->mi.bucket_size;
				}
				break;
			}
			}
		}
	}

	if (underflow_err.pos) {
		bool print = bch2_count_fsck_err(c, accounting_key_underflow, &underflow_err);
		unsigned pos = underflow_err.pos;
		int ret = bch2_run_explicit_recovery_pass(c, &underflow_err,
						      BCH_RECOVERY_PASS_check_allocations, 0);
		print |= underflow_err.pos != pos;

		if (print)
			bch2_print_str(c, KERN_ERR, underflow_err.buf);
		if (ret)
			return ret;
	}

	return 0;
}

/*
 * At startup time, initialize the in memory accounting from the btree (and
 * journal)
 */
int bch2_accounting_read(struct bch_fs *c)
{
	struct bch_accounting_mem *acc = &c->accounting;
	CLASS(btree_trans, trans)(c);
	CLASS(printbuf, buf)();

	/*
	 * We might run more than once if we rewind to start topology repair or
	 * btree node scan - and those might cause us to get different results,
	 * so we can't just skip if we've already run.
	 *
	 * Instead, zero out any accounting we have:
	 */
	scoped_guard(percpu_write, &c->mark_lock) {
		darray_for_each(acc->k, e)
			percpu_memset(e->v[0], 0, sizeof(u64) * e->nr_counters);
		for_each_member_device(c, ca)
			percpu_memset(ca->usage, 0, sizeof(*ca->usage));
		percpu_memset(c->usage, 0, sizeof(*c->usage));
	}

	struct journal_keys *keys = &c->journal_keys;
	struct journal_key *jk = keys->data;

	move_gap(keys, keys->nr);

	/* Find the range of accounting keys from the journal: */

	while (jk < &darray_top(*keys) &&
	       __journal_key_cmp(c, BTREE_ID_accounting, 0, POS_MIN, jk) > 0)
		jk++;

	struct journal_key *end = jk;
	while (end < &darray_top(*keys) &&
	       __journal_key_cmp(c, BTREE_ID_accounting, 0, SPOS_MAX, end) > 0)
		end++;

	/*
	 * Iterate over btree and journal accounting simultaneously:
	 *
	 * We want to drop unneeded unneeded accounting deltas early - deltas
	 * that are older than the btree accounting key version, and once we've
	 * dropped old accounting deltas we can accumulate and compact deltas
	 * for the same key:
	 */

	struct btree_iter iter;
	bch2_trans_iter_init(trans, &iter, BTREE_ID_accounting, POS_MIN,
			     BTREE_ITER_prefetch|BTREE_ITER_all_snapshots);
	iter.flags &= ~BTREE_ITER_with_journal;
	int ret = for_each_btree_key_continue(trans, iter,
				BTREE_ITER_prefetch|BTREE_ITER_all_snapshots, k, ({
		if (k.k->type != KEY_TYPE_accounting)
			continue;

		while (jk < end &&
		       __journal_key_cmp(c, BTREE_ID_accounting, 0, k.k->p, jk) > 0)
			jk = accumulate_and_read_journal_accounting(trans, jk);

		while (jk < end &&
		       __journal_key_cmp(c, BTREE_ID_accounting, 0, k.k->p, jk) == 0 &&
		       bversion_cmp(journal_key_k(c, jk)->k.bversion, k.k->bversion) <= 0) {
			jk->overwritten = true;
			jk++;
		}

		if (jk < end &&
		    __journal_key_cmp(c, BTREE_ID_accounting, 0, k.k->p, jk) == 0)
			jk = accumulate_and_read_journal_accounting(trans, jk);

		struct disk_accounting_pos acc_k;
		bpos_to_disk_accounting_pos(&acc_k, k.k->p);

		if (acc_k.type >= BCH_DISK_ACCOUNTING_TYPE_NR)
			break;

		if (!bch2_accounting_is_mem(&acc_k)) {
			struct disk_accounting_pos next_acc;
			memset(&next_acc, 0, sizeof(next_acc));
			next_acc.type = acc_k.type + 1;
			struct bpos next = disk_accounting_pos_to_bpos(&next_acc);
			if (jk < end)
				next = bpos_min(next, journal_key_k(c, jk)->k.p);

			/* for_each_btree_key() will still advance iterator: */
			bch2_btree_iter_set_pos(&iter, bpos_predecessor(next));
			continue;
		}

		accounting_read_key(trans, k);
	}));
	bch2_trans_iter_exit(&iter);
	if (ret)
		return ret;

	while (jk < end)
		jk = accumulate_and_read_journal_accounting(trans, jk);

	bch2_trans_unlock(trans);

	struct journal_key *dst = keys->data;
	darray_for_each(*keys, i)
		if (!i->overwritten)
			*dst++ = *i;
	keys->gap = keys->nr = dst - keys->data;

	return accounting_read_mem_fixups(trans);
}

int bch2_dev_usage_remove(struct bch_fs *c, struct bch_dev *ca)
{
	CLASS(btree_trans, trans)(c);

	struct disk_accounting_pos start;
	disk_accounting_key_init(start, dev_data_type, .dev = ca->dev_idx);

	struct disk_accounting_pos end;
	disk_accounting_key_init(end, dev_data_type, .dev = ca->dev_idx, .data_type = U8_MAX);

	return bch2_btree_write_buffer_flush_sync(trans) ?:
		commit_do(trans, NULL, NULL, 0, ({
			struct bkey_s_c k;
			int ret = 0;

			for_each_btree_key_max_norestart(trans, iter, BTREE_ID_accounting,
							 disk_accounting_pos_to_bpos(&start),
							 disk_accounting_pos_to_bpos(&end),
							 BTREE_ITER_all_snapshots, k, ret) {
				if (k.k->type != KEY_TYPE_accounting)
					continue;

				struct disk_accounting_pos acc;
				bpos_to_disk_accounting_pos(&acc, k.k->p);

				const unsigned nr = bch2_accounting_counters(k.k);
				u64 v[BCH_ACCOUNTING_MAX_COUNTERS];
				memcpy_u64s_small(v, bkey_s_c_to_accounting(k).v->d, nr);

				bch2_u64s_neg(v, nr);

				ret = bch2_disk_accounting_mod(trans, &acc, v, nr, false);
				if (ret)
					break;
			}

			ret;
	})) ?: bch2_btree_write_buffer_flush_sync(trans);
}

int bch2_dev_usage_init(struct bch_dev *ca, bool gc)
{
	struct bch_fs *c = ca->fs;
	CLASS(btree_trans, trans)(c);
	u64 v[3] = { ca->mi.nbuckets - ca->mi.first_bucket, 0, 0 };

	int ret = lockrestart_do(trans, ({
		bch2_disk_accounting_mod2(trans, gc,
					  v, dev_data_type,
					  .dev = ca->dev_idx,
					  .data_type = BCH_DATA_free) ?:
		(!gc ? bch2_trans_commit(trans, NULL, NULL, 0) : 0);
	}));
	bch_err_fn(c, ret);
	return ret;
}

void bch2_verify_accounting_clean(struct bch_fs *c)
{
	bool mismatch = false;
	struct bch_fs_usage_base base = {}, base_inmem = {};

	CLASS(btree_trans, trans)(c);
	for_each_btree_key(trans, iter,
			   BTREE_ID_accounting, POS_MIN,
			   BTREE_ITER_all_snapshots, k, ({
		u64 v[BCH_ACCOUNTING_MAX_COUNTERS];
		struct bkey_s_c_accounting a = bkey_s_c_to_accounting(k);
		unsigned nr = bch2_accounting_counters(k.k);

		struct disk_accounting_pos acc_k;
		bpos_to_disk_accounting_pos(&acc_k, k.k->p);

		if (acc_k.type >= BCH_DISK_ACCOUNTING_TYPE_NR)
			break;

		if (!bch2_accounting_is_mem(&acc_k)) {
			struct disk_accounting_pos next;
			memset(&next, 0, sizeof(next));
			next.type = acc_k.type + 1;
			bch2_btree_iter_set_pos(&iter, disk_accounting_pos_to_bpos(&next));
			continue;
		}

		bch2_accounting_mem_read(c, k.k->p, v, nr);

		if (memcmp(a.v->d, v, nr * sizeof(u64))) {
			CLASS(printbuf, buf)();

			bch2_bkey_val_to_text(&buf, c, k);
			prt_str(&buf, " !=");
			for (unsigned j = 0; j < nr; j++)
				prt_printf(&buf, " %llu", v[j]);

			pr_err("%s", buf.buf);
			mismatch = true;
		}

		switch (acc_k.type) {
		case BCH_DISK_ACCOUNTING_persistent_reserved:
			base.reserved += acc_k.persistent_reserved.nr_replicas * a.v->d[0];
			break;
		case BCH_DISK_ACCOUNTING_replicas:
			fs_usage_data_type_to_base(&base, acc_k.replicas.data_type, a.v->d[0]);
			break;
		case BCH_DISK_ACCOUNTING_dev_data_type: {
			{
				guard(rcu)(); /* scoped guard is a loop, and doesn't play nicely with continue */
				const enum bch_data_type data_type = acc_k.dev_data_type.data_type;
				struct bch_dev *ca = bch2_dev_rcu_noerror(c, acc_k.dev_data_type.dev);
				if (!ca)
					continue;

				v[0] = percpu_u64_get(&ca->usage->d[data_type].buckets);
				v[1] = percpu_u64_get(&ca->usage->d[data_type].sectors);
				v[2] = percpu_u64_get(&ca->usage->d[data_type].fragmented);

				if (data_type == BCH_DATA_sb || data_type == BCH_DATA_journal)
					base.hidden += a.v->d[0] * ca->mi.bucket_size;
			}

			if (memcmp(a.v->d, v, 3 * sizeof(u64))) {
				CLASS(printbuf, buf)();

				bch2_bkey_val_to_text(&buf, c, k);
				prt_str(&buf, " in mem");
				for (unsigned j = 0; j < nr; j++)
					prt_printf(&buf, " %llu", v[j]);

				pr_err("dev accounting mismatch: %s", buf.buf);
				mismatch = true;
			}
		}
		}

		0;
	}));

	acc_u64s_percpu(&base_inmem.hidden, &c->usage->hidden, sizeof(base_inmem) / sizeof(u64));

#define check(x)										\
	if (base.x != base_inmem.x) {								\
		pr_err("fs_usage_base.%s mismatch: %llu != %llu", #x, base.x, base_inmem.x);	\
		mismatch = true;								\
	}

	check(hidden);
	check(btree);
	check(data);
	check(cached);
	check(reserved);

	WARN_ON(mismatch);
}

void bch2_accounting_gc_free(struct bch_fs *c)
{
	lockdep_assert_held(&c->mark_lock);

	struct bch_accounting_mem *acc = &c->accounting;

	bch2_accounting_free_counters(acc, true);
	acc->gc_running = false;
}

void bch2_fs_accounting_exit(struct bch_fs *c)
{
	struct bch_accounting_mem *acc = &c->accounting;

	bch2_accounting_free_counters(acc, false);
	darray_exit(&acc->k);
}