linux/fs/bcachefs/movinggc.c

// SPDX-License-Identifier: GPL-2.0
/*
 * Moving/copying garbage collector
 *
 * Copyright 2012 Google, Inc.
 */

#include "bcachefs.h"
#include "alloc_foreground.h"
#include "btree_iter.h"
#include "btree_update.h"
#include "buckets.h"
#include "clock.h"
#include "disk_groups.h"
#include "error.h"
#include "extents.h"
#include "eytzinger.h"
#include "io.h"
#include "keylist.h"
#include "move.h"
#include "movinggc.h"
#include "super-io.h"
#include "trace.h"

#include <linux/freezer.h>
#include <linux/kthread.h>
#include <linux/math64.h>
#include <linux/sched/task.h>
#include <linux/sort.h>
#include <linux/wait.h>

/*
 * We can't use the entire copygc reserve in one iteration of copygc: we may
 * need the buckets we're freeing up to go back into the copygc reserve to make
 * forward progress, but if the copygc reserve is full they'll be available for
 * any allocation - and it's possible that in a given iteration, we free up most
 * of the buckets we're going to free before we allocate most of the buckets
 * we're going to allocate.
 *
 * If we only use half of the reserve per iteration, then in steady state we'll
 * always have room in the reserve for the buckets we're going to need in the
 * next iteration:
 */
#define COPYGC_BUCKETS_PER_ITER(ca)					\
	((ca)->free[RESERVE_MOVINGGC].size / 2)

static int bucket_offset_cmp(const void *_l, const void *_r, size_t size)
{
	const struct copygc_heap_entry *l = _l;
	const struct copygc_heap_entry *r = _r;

	return  cmp_int(l->dev,    r->dev) ?:
		cmp_int(l->offset, r->offset);
}

static enum data_cmd copygc_pred(struct bch_fs *c, void *arg,
				 struct bkey_s_c k,
				 struct bch_io_opts *io_opts,
				 struct data_opts *data_opts)
{
	copygc_heap *h = &c->copygc_heap;
	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
	const union bch_extent_entry *entry;
	struct extent_ptr_decoded p = { 0 };

	bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
		struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
		struct copygc_heap_entry search = {
			.dev	= p.ptr.dev,
			.offset	= p.ptr.offset,
		};

		ssize_t i = eytzinger0_find_le(h->data, h->used,
					       sizeof(h->data[0]),
					       bucket_offset_cmp, &search);
#if 0
		/* eytzinger search verify code: */
		ssize_t j = -1, k;

		for (k = 0; k < h->used; k++)
			if (h->data[k].offset <= ptr->offset &&
			    (j < 0 || h->data[k].offset > h->data[j].offset))
				j = k;

		BUG_ON(i != j);
#endif
		if (i >= 0 &&
		    p.ptr.offset < h->data[i].offset + ca->mi.bucket_size &&
		    p.ptr.gen == h->data[i].gen) {
			data_opts->target		= io_opts->background_target;
			data_opts->nr_replicas		= 1;
			data_opts->btree_insert_flags	= BTREE_INSERT_USE_RESERVE;
			data_opts->rewrite_dev		= p.ptr.dev;

			if (p.has_ec)
				data_opts->nr_replicas += p.ec.redundancy;

			return DATA_REWRITE;
		}
	}

	return DATA_SKIP;
}

static bool have_copygc_reserve(struct bch_dev *ca)
{
	bool ret;

	spin_lock(&ca->fs->freelist_lock);
	ret = fifo_full(&ca->free[RESERVE_MOVINGGC]) ||
		ca->allocator_state != ALLOCATOR_RUNNING;
	spin_unlock(&ca->fs->freelist_lock);

	return ret;
}

static inline int fragmentation_cmp(copygc_heap *heap,
				   struct copygc_heap_entry l,
				   struct copygc_heap_entry r)
{
	return cmp_int(l.fragmentation, r.fragmentation);
}

static int bch2_copygc(struct bch_fs *c)
{
	copygc_heap *h = &c->copygc_heap;
	struct copygc_heap_entry e, *i;
	struct bucket_array *buckets;
	struct bch_move_stats move_stats;
	u64 sectors_to_move = 0, sectors_not_moved = 0;
	u64 sectors_reserved = 0;
	u64 buckets_to_move, buckets_not_moved = 0;
	struct bch_dev *ca;
	unsigned dev_idx;
	size_t b, heap_size = 0;
	int ret;

	memset(&move_stats, 0, sizeof(move_stats));
	/*
	 * Find buckets with lowest sector counts, skipping completely
	 * empty buckets, by building a maxheap sorted by sector count,
	 * and repeatedly replacing the maximum element until all
	 * buckets have been visited.
	 */
	h->used = 0;

	for_each_rw_member(ca, c, dev_idx)
		heap_size += ca->mi.nbuckets >> 7;

	if (h->size < heap_size) {
		free_heap(&c->copygc_heap);
		if (!init_heap(&c->copygc_heap, heap_size, GFP_KERNEL)) {
			bch_err(c, "error allocating copygc heap");
			return 0;
		}
	}

	for_each_rw_member(ca, c, dev_idx) {
		closure_wait_event(&c->freelist_wait, have_copygc_reserve(ca));

		spin_lock(&ca->fs->freelist_lock);
		sectors_reserved += fifo_used(&ca->free[RESERVE_MOVINGGC]) * ca->mi.bucket_size;
		spin_unlock(&ca->fs->freelist_lock);

		down_read(&ca->bucket_lock);
		buckets = bucket_array(ca);

		for (b = buckets->first_bucket; b < buckets->nbuckets; b++) {
			struct bucket *g = buckets->b + b;
			struct bucket_mark m = READ_ONCE(g->mark);
			struct copygc_heap_entry e;

			if (m.owned_by_allocator ||
			    m.data_type != BCH_DATA_user ||
			    !bucket_sectors_used(m) ||
			    bucket_sectors_used(m) >= ca->mi.bucket_size)
				continue;

			WARN_ON(m.stripe && !g->stripe_redundancy);

			e = (struct copygc_heap_entry) {
				.dev		= dev_idx,
				.gen		= m.gen,
				.replicas	= 1 + g->stripe_redundancy,
				.fragmentation	= bucket_sectors_used(m) * (1U << 15)
					/ ca->mi.bucket_size,
				.sectors	= bucket_sectors_used(m),
				.offset		= bucket_to_sector(ca, b),
			};
			heap_add_or_replace(h, e, -fragmentation_cmp, NULL);
		}
		up_read(&ca->bucket_lock);
	}

	if (!sectors_reserved) {
		bch2_fs_fatal_error(c, "stuck, ran out of copygc reserve!");
		return -1;
	}

	/*
	 * Our btree node allocations also come out of RESERVE_MOVINGGC:
	 */
	sectors_to_move = (sectors_to_move * 3) / 4;

	for (i = h->data; i < h->data + h->used; i++)
		sectors_to_move += i->sectors * i->replicas;

	while (sectors_to_move > sectors_reserved) {
		BUG_ON(!heap_pop(h, e, -fragmentation_cmp, NULL));
		sectors_to_move -= e.sectors * e.replicas;
	}

	buckets_to_move = h->used;

	if (!buckets_to_move)
		return 0;

	eytzinger0_sort(h->data, h->used,
			sizeof(h->data[0]),
			bucket_offset_cmp, NULL);

	ret = bch2_move_data(c, &c->copygc_pd.rate,
			     writepoint_ptr(&c->copygc_write_point),
			     POS_MIN, POS_MAX,
			     copygc_pred, NULL,
			     &move_stats);

	for_each_rw_member(ca, c, dev_idx) {
		down_read(&ca->bucket_lock);
		buckets = bucket_array(ca);
		for (i = h->data; i < h->data + h->used; i++) {
			struct bucket_mark m;
			size_t b;

			if (i->dev != dev_idx)
				continue;

			b = sector_to_bucket(ca, i->offset);
			m = READ_ONCE(buckets->b[b].mark);

			if (i->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_ratelimited(c,
			"copygc finished but %llu/%llu sectors, %llu/%llu buckets not moved (move stats: moved %llu sectors, raced %llu keys, %llu sectors)",
			 sectors_not_moved, sectors_to_move,
			 buckets_not_moved, buckets_to_move,
			 atomic64_read(&move_stats.sectors_moved),
			 atomic64_read(&move_stats.keys_raced),
			 atomic64_read(&move_stats.sectors_raced));

	trace_copygc(c,
		     atomic64_read(&move_stats.sectors_moved), sectors_not_moved,
		     buckets_to_move, buckets_not_moved);
	return 0;
}

/*
 * Copygc runs when the amount of fragmented data is above some arbitrary
 * threshold:
 *
 * The threshold at the limit - when the device is full - is the amount of space
 * we reserved in bch2_recalc_capacity; we can't have more than that amount of
 * disk space stranded due to fragmentation and store everything we have
 * promised to store.
 *
 * But we don't want to be running copygc unnecessarily when the device still
 * has plenty of free space - rather, we want copygc to smoothly run every so
 * often and continually reduce the amount of fragmented space as the device
 * fills up. So, we increase the threshold by half the current free space.
 */
unsigned long bch2_copygc_wait_amount(struct bch_fs *c)
{
	struct bch_dev *ca;
	unsigned dev_idx;
	u64 fragmented_allowed = c->copygc_threshold;
	u64 fragmented = 0;

	for_each_rw_member(ca, c, dev_idx) {
		struct bch_dev_usage usage = bch2_dev_usage_read(ca);

		fragmented_allowed += ((__dev_buckets_available(ca, usage) *
					ca->mi.bucket_size) >> 1);
		fragmented += usage.d[BCH_DATA_user].fragmented;
	}

	return max_t(s64, 0, fragmented_allowed - fragmented);
}

static int bch2_copygc_thread(void *arg)
{
	struct bch_fs *c = arg;
	struct io_clock *clock = &c->io_clock[WRITE];
	u64 last, wait;

	set_freezable();

	while (!kthread_should_stop()) {
		if (kthread_wait_freezable(c->copy_gc_enabled))
			break;

		last = atomic64_read(&clock->now);
		wait = bch2_copygc_wait_amount(c);

		if (wait > clock->max_slop) {
			bch2_kthread_io_clock_wait(clock, last + wait,
					MAX_SCHEDULE_TIMEOUT);
			continue;
		}

		if (bch2_copygc(c))
			break;
	}

	return 0;
}

void bch2_copygc_stop(struct bch_fs *c)
{
	c->copygc_pd.rate.rate = UINT_MAX;
	bch2_ratelimit_reset(&c->copygc_pd.rate);

	if (c->copygc_thread) {
		kthread_stop(c->copygc_thread);
		put_task_struct(c->copygc_thread);
	}
	c->copygc_thread = NULL;
}

int bch2_copygc_start(struct bch_fs *c)
{
	struct task_struct *t;

	if (c->copygc_thread)
		return 0;

	if (c->opts.nochanges)
		return 0;

	if (bch2_fs_init_fault("copygc_start"))
		return -ENOMEM;

	t = kthread_create(bch2_copygc_thread, c, "bch-copygc/%s", c->name);
	if (IS_ERR(t))
		return PTR_ERR(t);

	get_task_struct(t);

	c->copygc_thread = t;
	wake_up_process(c->copygc_thread);

	return 0;
}

void bch2_fs_copygc_init(struct bch_fs *c)
{
	bch2_pd_controller_init(&c->copygc_pd);
	c->copygc_pd.d_term = 0;
}