On Mon, 8 Apr 2024, Coly Li wrote:
> On Sun, Apr 07, 2024 at 03:54:30PM -0700, Eric Wheeler wrote:
> > Hi Coly,
> >
>
> Hi Eric,
>
> > Did you see the patches from Mingzhe from July 19th? They look like good
> > ideas but I've not seen any traction or commentary yet. Below they are
> > forwarded for your reference.
> >
> >
>
> Yes I knew this series. This is a performance optimization with a bit
> complicated change. The overall gain is not that easy to esitamte.
>
> For such complicated optimization, I am always quite conservative and
> careful. Because in many cases, another performance regression introduced
> with a specific performance gain observed. Therefore, if the coding
> itself is good, I'd like to say long term result before picking it.
>
> During Sep 2023, there was a meetup in Shenzhen called "China LSF",
> which was for local developers to meet and communicate, which was quite
> helpful for people who didn't have fund for oversea travel or preferred
> to talk in mandarin. I met Mingzhe face to face during CLSF2023, and
> discussed this series.
Fun!
> I mentioned the counter related stuffs should be improved, and I also
> suggest to use it in product environment to see whether real performance
> gain can be obviously observed.
>
> From later update from Mingzhe, the expected performance improvement was
> not significantly observed, and not adopted in their product neither.
Ok thanks for explaining, I understand.
> So far, another proactive bucket invalidate patch (from Mingzhe) looks
> fine to me and I use it in my following performance testing now. But for
> the series you mentioned in this email thread, there is little chance
> to have it in mainline and I just leave it in silence.
>
> BTW, do you observe stable performance improvement from this series?
I've not tried the patch series because I was waiting for list feedback
before trying it out. It wasn't performance that caught my eye: While
performance may be a side-effect, I was interested in these for the
following reasons:
* bcache: only copy dirty data during moving gc
For those who burn through the TBW/DWPD disks writes quickly,
reducing write cycles is a win because it saves people on real
hardware costs, plus the value of reduced environmental impact due
to failed/worn out SSDs. Mingzhe wrote:
"Usually ssd and nvme calculate the lifespan by the write cycles.
When moving gc, only copy dirty data, which can reduce the amount
of written data. This will improve moving gc speed, and extend
cache life."
* bcache: the gc_sectors_used size matches the bucket size
If I understand correctly, this looks like a good sanity check,
and maybe it should cc stable@?
* bcache: Separate bch_moving_gc() from bch_btree_gc()
Latency is critical for most applications using bcache, and this
prevents IO request blocking unless the fragments reach a
threshold.
--
Eric Wheeler
> Thanks.
>
> Coly Li
>
>
> > --
> > Eric Wheeler
> >
> > ---------- Forwarded message ----------
> > In-Reply-To: <20230719072753.366-1-mingzhe.zou@xxxxxxxxxxxx>
> > Date: Wed, 19 Jul 2023 15:27:52 +0800
> > To: colyli@xxxxxxx, linux-bcache@xxxxxxxxxxxxxxx
> > Cc: bcache@xxxxxxxxxxxxxxxxxx, zoumingzhe@xxxxxx
> > From: Mingzhe Zou <mingzhe.zou@xxxxxxxxxxxx>
> > Subject: [PATCH v3 2/3] bcache: Separate bch_moving_gc() from bch_btree_gc()
> >
> > From: Mingzhe Zou <zoumingzhe@xxxxxx>
> >
> > Moving gc uses cache->heap to defragment disk. Unlike btree gc,
> > moving gc only takes up part of the disk bandwidth.
> >
> > The number of heap is constant. However, the buckets released by
> > each moving gc is limited. So bch_moving_gc() needs to be called
> > multiple times.
> >
> > If bch_gc_thread() always calls bch_btree_gc(), it will block
> > the IO request.This patch allows bch_gc_thread() to only call
> > bch_moving_gc() when there are many fragments.
> >
> > Signed-off-by: Mingzhe Zou <mingzhe.zou@xxxxxxxxxxxx>
> > ---
> > drivers/md/bcache/bcache.h | 4 +-
> > drivers/md/bcache/btree.c | 73 ++++++++++++++++++++++++++++++++++--
> > drivers/md/bcache/movinggc.c | 7 +++-
> > drivers/md/bcache/super.c | 2 +
> > 4 files changed, 81 insertions(+), 5 deletions(-)
> >
> > diff --git a/drivers/md/bcache/bcache.h b/drivers/md/bcache/bcache.h
> > index 582df3c9dc1b..10f3f548629e 100644
> > --- a/drivers/md/bcache/bcache.h
> > +++ b/drivers/md/bcache/bcache.h
> > @@ -459,7 +459,8 @@ struct cache {
> > * until a gc finishes - otherwise we could pointlessly burn a ton of
> > * cpu
> > */
> > - unsigned int invalidate_needs_gc;
> > + unsigned int invalidate_needs_gc:1;
> > + unsigned int only_moving_gc:1;
> >
> > bool discard; /* Get rid of? */
> >
> > @@ -627,6 +628,7 @@ struct cache_set {
> > struct gc_stat gc_stats;
> > size_t nbuckets;
> > size_t avail_nbuckets;
> > + size_t fragment_nbuckets;
> >
> > struct task_struct *gc_thread;
> > /* Where in the btree gc currently is */
> > diff --git a/drivers/md/bcache/btree.c b/drivers/md/bcache/btree.c
> > index ca962f329977..baa2149e9235 100644
> > --- a/drivers/md/bcache/btree.c
> > +++ b/drivers/md/bcache/btree.c
> > @@ -88,6 +88,7 @@
> > * Test module load/unload
> > */
> >
> > +#define COPY_GC_PERCENT 5
> > #define MAX_NEED_GC 64
> > #define MAX_SAVE_PRIO 72
> > #define MAX_GC_TIMES 100
> > @@ -1726,6 +1727,7 @@ static void btree_gc_start(struct cache_set *c)
> >
> > mutex_lock(&c->bucket_lock);
> >
> > + set_gc_sectors(c);
> > c->gc_mark_valid = 0;
> > c->gc_done = ZERO_KEY;
> >
> > @@ -1852,8 +1854,58 @@ static void bch_btree_gc(struct cache_set *c)
> > memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat));
> >
> > trace_bcache_gc_end(c);
> > +}
> > +
> > +extern unsigned int bch_cutoff_writeback;
> > +extern unsigned int bch_cutoff_writeback_sync;
> > +
> > +static bool moving_gc_should_run(struct cache_set *c)
> > +{
> > + struct bucket *b;
> > + struct cache *ca = c->cache;
> > + size_t moving_gc_threshold = ca->sb.bucket_size >> 2, frag_percent;
> > + unsigned long used_buckets = 0, frag_buckets = 0, move_buckets = 0;
> > + unsigned long dirty_sectors = 0, frag_sectors = 0, used_sectors = 0;
> > +
> > + mutex_lock(&c->bucket_lock);
> > + for_each_bucket(b, ca) {
> > + if (GC_MOVE(b) || GC_MARK(b) != GC_MARK_DIRTY)
> > + continue;
> > +
> > + used_buckets++;
> > +
> > + used_sectors = b->gc_sectors_used;
> > + dirty_sectors += used_sectors;
> > +
> > + if (used_sectors < ca->sb.bucket_size)
> > + frag_buckets++;
> > +
> > + if (used_sectors <= moving_gc_threshold)
> > + move_buckets++;
> > + }
> > + mutex_unlock(&c->bucket_lock);
> > +
> > + c->fragment_nbuckets = frag_buckets;
> >
> > - bch_moving_gc(c);
> > + if (used_buckets < c->nbuckets * bch_cutoff_writeback / 100)
> > + return false;
> > +
> > + if (move_buckets > ca->heap.size)
> > + return true;
> > +
> > + frag_sectors = used_buckets * ca->sb.bucket_size - dirty_sectors;
> > + frag_percent = div_u64(frag_sectors * 100, ca->sb.bucket_size * c->nbuckets);
> > +
> > + if (frag_percent >= COPY_GC_PERCENT)
> > + return true;
> > +
> > + if (used_buckets > c->nbuckets * bch_cutoff_writeback_sync / 100)
> > + return true;
> > +
> > + if (c->gc_stats.in_use > bch_cutoff_writeback_sync && frag_buckets > 0)
> > + return true;
> > +
> > + return false;
> > }
> >
> > static bool gc_should_run(struct cache_set *c)
> > @@ -1866,6 +1918,19 @@ static bool gc_should_run(struct cache_set *c)
> > if (atomic_read(&c->sectors_to_gc) < 0)
> > return true;
> >
> > + /*
> > + * Moving gc uses cache->heap to defragment disk. Unlike btree gc,
> > + * moving gc only takes up part of the disk bandwidth.
> > + * The number of heap is constant. However, the buckets released by
> > + * each moving gc is limited. So bch_moving_gc() needs to be called
> > + * multiple times. If bch_gc_thread() always calls bch_btree_gc(),
> > + * it will block the IO request.
> > + */
> > + if (c->copy_gc_enabled && moving_gc_should_run(c)) {
> > + ca->only_moving_gc = 1;
> > + return true;
> > + }
> > +
> > return false;
> > }
> >
> > @@ -1883,8 +1948,10 @@ static int bch_gc_thread(void *arg)
> > test_bit(CACHE_SET_IO_DISABLE, &c->flags))
> > break;
> >
> > - set_gc_sectors(c);
> > - bch_btree_gc(c);
> > + if (!c->cache->only_moving_gc)
> > + bch_btree_gc(c);
> > +
> > + bch_moving_gc(c);
> > }
> >
> > wait_for_kthread_stop();
> > diff --git a/drivers/md/bcache/movinggc.c b/drivers/md/bcache/movinggc.c
> > index e4182c3ba9f8..2c3cb1b6e0af 100644
> > --- a/drivers/md/bcache/movinggc.c
> > +++ b/drivers/md/bcache/movinggc.c
> > @@ -200,6 +200,8 @@ void bch_moving_gc(struct cache_set *c)
> > struct bucket *b;
> > unsigned long sectors_to_move, reserve_sectors;
> >
> > + c->cache->only_moving_gc = 0;
> > +
> > if (!c->copy_gc_enabled)
> > return;
> >
> > @@ -212,7 +214,7 @@ void bch_moving_gc(struct cache_set *c)
> > ca->heap.used = 0;
> >
> > for_each_bucket(b, ca) {
> > - if (GC_MARK(b) == GC_MARK_METADATA ||
> > + if (GC_MOVE(b) || GC_MARK(b) == GC_MARK_METADATA ||
> > !b->gc_sectors_used ||
> > b->gc_sectors_used == ca->sb.bucket_size ||
> > atomic_read(&b->pin))
> > @@ -229,6 +231,9 @@ void bch_moving_gc(struct cache_set *c)
> > heap_sift(&ca->heap, 0, bucket_cmp);
> > }
> > }
> > +
> > + pr_info("moving gc: on set %pU, %lu sectors from %zu buckets",
> > + c->set_uuid, sectors_to_move, ca->heap.used);
> >
> > while (sectors_to_move > reserve_sectors) {
> > heap_pop(&ca->heap, b, bucket_cmp);
> > diff --git a/drivers/md/bcache/super.c b/drivers/md/bcache/super.c
> > index 0ae2b3676293..7e556bc0ec04 100644
> > --- a/drivers/md/bcache/super.c
> > +++ b/drivers/md/bcache/super.c
> > @@ -2112,6 +2112,8 @@ static int run_cache_set(struct cache_set *c)
> > if (bch_gc_thread_start(c))
> > goto err;
> >
> > + force_wake_up_gc(c);
> > +
> > closure_sync(&cl);
> > c->cache->sb.last_mount = (u32)ktime_get_real_seconds();
> > bcache_write_super(c);
> > --
> > 2.17.1.windows.2
> >
> >
> >
> >
> > ---------- Forwarded message ----------
> > Date: Wed, 19 Jul 2023 15:27:51 +0800
> > To: colyli@xxxxxxx, linux-bcache@xxxxxxxxxxxxxxx
> > Cc: bcache@xxxxxxxxxxxxxxxxxx, zoumingzhe@xxxxxx
> > From: Mingzhe Zou <mingzhe.zou@xxxxxxxxxxxx>
> > Subject: [PATCH v3 1/3] bcache: the gc_sectors_used size matches the bucket size
> >
> > From: Mingzhe Zou <zoumingzhe@xxxxxx>
> >
> > The bucket size in the superblock is defined as uint16_t.
> > But, GC_SECTORS_USED is only 13 bits. If the bucket size
> > is 4M (8192 sectors), GC_SECTORS_USED will be truncated
> > to MAX_GC_SECTORS_USED.
> >
> > GC_SECTORS_USED is the moving gc sorting condition, we
> > should try our best to ensure it is correct.
> >
> > Signed-off-by: Mingzhe Zou <mingzhe.zou@xxxxxxxxxxxx>
> > ---
> > drivers/md/bcache/alloc.c | 12 ++++++++----
> > drivers/md/bcache/bcache.h | 12 ++++++------
> > drivers/md/bcache/btree.c | 18 ++++++++++++------
> > drivers/md/bcache/movinggc.c | 14 +++++++-------
> > drivers/md/bcache/sysfs.c | 2 +-
> > include/trace/events/bcache.h | 2 +-
> > 6 files changed, 35 insertions(+), 25 deletions(-)
> >
> > diff --git a/drivers/md/bcache/alloc.c b/drivers/md/bcache/alloc.c
> > index ce13c272c387..4ae1018bf029 100644
> > --- a/drivers/md/bcache/alloc.c
> > +++ b/drivers/md/bcache/alloc.c
> > @@ -142,7 +142,7 @@ void __bch_invalidate_one_bucket(struct cache *ca, struct bucket *b)
> > lockdep_assert_held(&ca->set->bucket_lock);
> > BUG_ON(GC_MARK(b) && GC_MARK(b) != GC_MARK_RECLAIMABLE);
> >
> > - if (GC_SECTORS_USED(b))
> > + if (b->gc_sectors_used)
> > trace_bcache_invalidate(ca, b - ca->buckets);
> >
> > bch_inc_gen(ca, b);
> > @@ -170,7 +170,7 @@ static void bch_invalidate_one_bucket(struct cache *ca, struct bucket *b)
> > ({ \
> > unsigned int min_prio = (INITIAL_PRIO - ca->set->min_prio) / 8; \
> > \
> > - (b->prio - ca->set->min_prio + min_prio) * GC_SECTORS_USED(b); \
> > + (b->prio - ca->set->min_prio + min_prio) * b->gc_sectors_used; \
> > })
> >
> > #define bucket_max_cmp(l, r) (bucket_prio(l) < bucket_prio(r))
> > @@ -446,7 +446,11 @@ long bch_bucket_alloc(struct cache *ca, unsigned int reserve, bool wait)
> >
> > BUG_ON(atomic_read(&b->pin) != 1);
> >
> > - SET_GC_SECTORS_USED(b, ca->sb.bucket_size);
> > + /*
> > + * If gc_sectors_used is 0, moving gc is preferred. But the new bucket
> > + * should not do moving gc. So we set gc_sectors_used to the maximum.
> > + */
> > + b->gc_sectors_used = ca->sb.bucket_size;
> >
> > if (reserve <= RESERVE_PRIO) {
> > SET_GC_MARK(b, GC_MARK_METADATA);
> > @@ -469,7 +473,7 @@ long bch_bucket_alloc(struct cache *ca, unsigned int reserve, bool wait)
> > void __bch_bucket_free(struct cache *ca, struct bucket *b)
> > {
> > SET_GC_MARK(b, 0);
> > - SET_GC_SECTORS_USED(b, 0);
> > + b->gc_sectors_used = 0;
> >
> > if (ca->set->avail_nbuckets < ca->set->nbuckets) {
> > ca->set->avail_nbuckets++;
> > diff --git a/drivers/md/bcache/bcache.h b/drivers/md/bcache/bcache.h
> > index 5a79bb3c272f..582df3c9dc1b 100644
> > --- a/drivers/md/bcache/bcache.h
> > +++ b/drivers/md/bcache/bcache.h
> > @@ -197,9 +197,10 @@
> > struct bucket {
> > atomic_t pin;
> > uint16_t prio;
> > + uint16_t flag;
> > uint8_t gen;
> > uint8_t last_gc; /* Most out of date gen in the btree */
> > - uint16_t gc_mark; /* Bitfield used by GC. See below for field */
> > + uint16_t gc_sectors_used;
> > };
> >
> > /*
> > @@ -207,14 +208,11 @@ struct bucket {
> > * as multiple threads touch struct bucket without locking
> > */
> >
> > -BITMASK(GC_MARK, struct bucket, gc_mark, 0, 2);
> > #define GC_MARK_RECLAIMABLE 1
> > #define GC_MARK_DIRTY 2
> > #define GC_MARK_METADATA 3
> > -#define GC_SECTORS_USED_SIZE 13
> > -#define MAX_GC_SECTORS_USED (~(~0ULL << GC_SECTORS_USED_SIZE))
> > -BITMASK(GC_SECTORS_USED, struct bucket, gc_mark, 2, GC_SECTORS_USED_SIZE);
> > -BITMASK(GC_MOVE, struct bucket, gc_mark, 15, 1);
> > +BITMASK(GC_MARK, struct bucket, flag, 0, 2);
> > +BITMASK(GC_MOVE, struct bucket, flag, 2, 1);
> >
> > #include "journal.h"
> > #include "stats.h"
> > @@ -764,6 +762,8 @@ struct bbio {
> > #define bucket_bytes(ca) ((ca)->sb.bucket_size << 9)
> > #define block_bytes(ca) ((ca)->sb.block_size << 9)
> >
> > +#define MAX_BUCKET_SIZE (~(~0ULL << 16)) /* sectors */
> > +
> > static inline unsigned int meta_bucket_pages(struct cache_sb *sb)
> > {
> > unsigned int n, max_pages;
> > diff --git a/drivers/md/bcache/btree.c b/drivers/md/bcache/btree.c
> > index fd121a61f17c..ca962f329977 100644
> > --- a/drivers/md/bcache/btree.c
> > +++ b/drivers/md/bcache/btree.c
> > @@ -1255,11 +1255,11 @@ static uint8_t __bch_btree_mark_key(struct cache_set *c, int level,
> > SET_GC_MARK(g, GC_MARK_RECLAIMABLE);
> >
> > /* guard against overflow */
> > - SET_GC_SECTORS_USED(g, min_t(unsigned int,
> > - GC_SECTORS_USED(g) + KEY_SIZE(k),
> > - MAX_GC_SECTORS_USED));
> > + g->gc_sectors_used = min_t(uint16_t, c->cache->sb.bucket_size,
> > + g->gc_sectors_used + KEY_SIZE(k));
> >
> > - BUG_ON(!GC_SECTORS_USED(g));
> > + BUG_ON(g->gc_sectors_used < KEY_SIZE(k) ||
> > + g->gc_sectors_used > c->cache->sb.bucket_size);
> > }
> >
> > return stale;
> > @@ -1732,9 +1732,15 @@ static void btree_gc_start(struct cache_set *c)
> > ca = c->cache;
> > for_each_bucket(b, ca) {
> > b->last_gc = b->gen;
> > +
> > + /*
> > + * If the bucket is still in use, mark is not necessary.
> > + * In bch_bucket_alloc(), we set the gc_sectors_used to
> > + * cache bucket size, just keep the maximum.
> > + */
> > if (!atomic_read(&b->pin)) {
> > SET_GC_MARK(b, 0);
> > - SET_GC_SECTORS_USED(b, 0);
> > + b->gc_sectors_used = 0;
> > }
> > }
> >
> > @@ -1797,7 +1803,7 @@ static void bch_btree_gc_finish(struct cache_set *c)
> > if (atomic_read(&b->pin))
> > continue;
> >
> > - BUG_ON(!GC_MARK(b) && GC_SECTORS_USED(b));
> > + BUG_ON(!GC_MARK(b) && b->gc_sectors_used);
> >
> > if (!GC_MARK(b) || GC_MARK(b) == GC_MARK_RECLAIMABLE)
> > c->avail_nbuckets++;
> > diff --git a/drivers/md/bcache/movinggc.c b/drivers/md/bcache/movinggc.c
> > index 9f32901fdad1..e4182c3ba9f8 100644
> > --- a/drivers/md/bcache/movinggc.c
> > +++ b/drivers/md/bcache/movinggc.c
> > @@ -184,14 +184,14 @@ err: if (!IS_ERR_OR_NULL(w->private))
> >
> > static bool bucket_cmp(struct bucket *l, struct bucket *r)
> > {
> > - return GC_SECTORS_USED(l) < GC_SECTORS_USED(r);
> > + return l->gc_sectors_used < r->gc_sectors_used;
> > }
> >
> > static unsigned int bucket_heap_top(struct cache *ca)
> > {
> > struct bucket *b;
> >
> > - return (b = heap_peek(&ca->heap)) ? GC_SECTORS_USED(b) : 0;
> > + return (b = heap_peek(&ca->heap)) ? b->gc_sectors_used : 0;
> > }
> >
> > void bch_moving_gc(struct cache_set *c)
> > @@ -213,17 +213,17 @@ void bch_moving_gc(struct cache_set *c)
> >
> > for_each_bucket(b, ca) {
> > if (GC_MARK(b) == GC_MARK_METADATA ||
> > - !GC_SECTORS_USED(b) ||
> > - GC_SECTORS_USED(b) == ca->sb.bucket_size ||
> > + !b->gc_sectors_used ||
> > + b->gc_sectors_used == ca->sb.bucket_size ||
> > atomic_read(&b->pin))
> > continue;
> >
> > if (!heap_full(&ca->heap)) {
> > - sectors_to_move += GC_SECTORS_USED(b);
> > + sectors_to_move += b->gc_sectors_used;
> > heap_add(&ca->heap, b, bucket_cmp);
> > } else if (bucket_cmp(b, heap_peek(&ca->heap))) {
> > sectors_to_move -= bucket_heap_top(ca);
> > - sectors_to_move += GC_SECTORS_USED(b);
> > + sectors_to_move += b->gc_sectors_used;
> >
> > ca->heap.data[0] = b;
> > heap_sift(&ca->heap, 0, bucket_cmp);
> > @@ -232,7 +232,7 @@ void bch_moving_gc(struct cache_set *c)
> >
> > while (sectors_to_move > reserve_sectors) {
> > heap_pop(&ca->heap, b, bucket_cmp);
> > - sectors_to_move -= GC_SECTORS_USED(b);
> > + sectors_to_move -= b->gc_sectors_used;
> > }
> >
> > while (heap_pop(&ca->heap, b, bucket_cmp))
> > diff --git a/drivers/md/bcache/sysfs.c b/drivers/md/bcache/sysfs.c
> > index 0e2c1880f60b..3b859954b8c5 100644
> > --- a/drivers/md/bcache/sysfs.c
> > +++ b/drivers/md/bcache/sysfs.c
> > @@ -1073,7 +1073,7 @@ SHOW(__bch_cache)
> >
> > mutex_lock(&ca->set->bucket_lock);
> > for_each_bucket(b, ca) {
> > - if (!GC_SECTORS_USED(b))
> > + if (!b->gc_sectors_used)
> > unused++;
> > if (GC_MARK(b) == GC_MARK_RECLAIMABLE)
> > available++;
> > diff --git a/include/trace/events/bcache.h b/include/trace/events/bcache.h
> > index 899fdacf57b9..b9d63e18c453 100644
> > --- a/include/trace/events/bcache.h
> > +++ b/include/trace/events/bcache.h
> > @@ -438,7 +438,7 @@ TRACE_EVENT(bcache_invalidate,
> > TP_fast_assign(
> > __entry->dev = ca->bdev->bd_dev;
> > __entry->offset = bucket << ca->set->bucket_bits;
> > - __entry->sectors = GC_SECTORS_USED(&ca->buckets[bucket]);
> > + __entry->sectors = ca->buckets[bucket].gc_sectors_used;
> > ),
> >
> > TP_printk("invalidated %u sectors at %d,%d sector=%llu",
> > --
> > 2.17.1.windows.2
> >
> >
> >
> >
> > ---------- Forwarded message ----------
> > In-Reply-To: <20230719072753.366-1-mingzhe.zou@xxxxxxxxxxxx>
> > Date: Wed, 19 Jul 2023 15:27:53 +0800
> > To: colyli@xxxxxxx, linux-bcache@xxxxxxxxxxxxxxx
> > Cc: bcache@xxxxxxxxxxxxxxxxxx, zoumingzhe@xxxxxx
> > From: Mingzhe Zou <mingzhe.zou@xxxxxxxxxxxx>
> > Subject: [PATCH v3 3/3] bcache: only copy dirty data during moving gc
> >
> > From: Mingzhe Zou <zoumingzhe@xxxxxx>
> >
> > When we want to shorten the moving gc interval, we must consider
> > its impact, such as: performance, cache life.
> >
> > Usually ssd and nvme calculate the lifespan by the write cycles.
> > When moving gc, only copy dirty data, which can reduce the amount
> > of written data. This will improve moving gc speed, and extend
> > cache life.
> >
> > Signed-off-by: Mingzhe Zou <mingzhe.zou@xxxxxxxxxxxx>
> > ---
> > drivers/md/bcache/alloc.c | 2 ++
> > drivers/md/bcache/bcache.h | 1 +
> > drivers/md/bcache/btree.c | 10 +++++++++-
> > drivers/md/bcache/movinggc.c | 16 ++++++++--------
> > 4 files changed, 20 insertions(+), 9 deletions(-)
> >
> > diff --git a/drivers/md/bcache/alloc.c b/drivers/md/bcache/alloc.c
> > index 4ae1018bf029..3d4b9f50b056 100644
> > --- a/drivers/md/bcache/alloc.c
> > +++ b/drivers/md/bcache/alloc.c
> > @@ -451,6 +451,7 @@ long bch_bucket_alloc(struct cache *ca, unsigned int reserve, bool wait)
> > * should not do moving gc. So we set gc_sectors_used to the maximum.
> > */
> > b->gc_sectors_used = ca->sb.bucket_size;
> > + b->gc_sectors_dirty = ca->sb.bucket_size;
> >
> > if (reserve <= RESERVE_PRIO) {
> > SET_GC_MARK(b, GC_MARK_METADATA);
> > @@ -474,6 +475,7 @@ void __bch_bucket_free(struct cache *ca, struct bucket *b)
> > {
> > SET_GC_MARK(b, 0);
> > b->gc_sectors_used = 0;
> > + b->gc_sectors_dirty = 0;
> >
> > if (ca->set->avail_nbuckets < ca->set->nbuckets) {
> > ca->set->avail_nbuckets++;
> > diff --git a/drivers/md/bcache/bcache.h b/drivers/md/bcache/bcache.h
> > index 10f3f548629e..b3e8e4f513f1 100644
> > --- a/drivers/md/bcache/bcache.h
> > +++ b/drivers/md/bcache/bcache.h
> > @@ -201,6 +201,7 @@ struct bucket {
> > uint8_t gen;
> > uint8_t last_gc; /* Most out of date gen in the btree */
> > uint16_t gc_sectors_used;
> > + uint16_t gc_sectors_dirty;
> > };
> >
> > /*
> > diff --git a/drivers/md/bcache/btree.c b/drivers/md/bcache/btree.c
> > index baa2149e9235..d4aeaaf1b9bc 100644
> > --- a/drivers/md/bcache/btree.c
> > +++ b/drivers/md/bcache/btree.c
> > @@ -1250,8 +1250,14 @@ static uint8_t __bch_btree_mark_key(struct cache_set *c, int level,
> >
> > if (level)
> > SET_GC_MARK(g, GC_MARK_METADATA);
> > - else if (KEY_DIRTY(k))
> > + else if (KEY_DIRTY(k)) {
> > SET_GC_MARK(g, GC_MARK_DIRTY);
> > + g->gc_sectors_dirty = min_t(uint16_t, c->cache->sb.bucket_size,
> > + g->gc_sectors_dirty + KEY_SIZE(k));
> > +
> > + BUG_ON(g->gc_sectors_dirty < KEY_SIZE(k) ||
> > + g->gc_sectors_dirty > c->cache->sb.bucket_size);
> > + }
> > else if (!GC_MARK(g))
> > SET_GC_MARK(g, GC_MARK_RECLAIMABLE);
> >
> > @@ -1743,6 +1749,7 @@ static void btree_gc_start(struct cache_set *c)
> > if (!atomic_read(&b->pin)) {
> > SET_GC_MARK(b, 0);
> > b->gc_sectors_used = 0;
> > + b->gc_sectors_dirty = 0;
> > }
> > }
> >
> > @@ -1806,6 +1813,7 @@ static void bch_btree_gc_finish(struct cache_set *c)
> > continue;
> >
> > BUG_ON(!GC_MARK(b) && b->gc_sectors_used);
> > + BUG_ON(!GC_MARK(b) && b->gc_sectors_dirty);
> >
> > if (!GC_MARK(b) || GC_MARK(b) == GC_MARK_RECLAIMABLE)
> > c->avail_nbuckets++;
> > diff --git a/drivers/md/bcache/movinggc.c b/drivers/md/bcache/movinggc.c
> > index 2c3cb1b6e0af..98f2487f518d 100644
> > --- a/drivers/md/bcache/movinggc.c
> > +++ b/drivers/md/bcache/movinggc.c
> > @@ -27,7 +27,7 @@ static bool moving_pred(struct keybuf *buf, struct bkey *k)
> >
> > for (i = 0; i < KEY_PTRS(k); i++)
> > if (ptr_available(c, k, i) &&
> > - GC_MOVE(PTR_BUCKET(c, k, i)))
> > + GC_MOVE(PTR_BUCKET(c, k, i)) && KEY_DIRTY(k))
> > return true;
> >
> > return false;
> > @@ -184,14 +184,14 @@ err: if (!IS_ERR_OR_NULL(w->private))
> >
> > static bool bucket_cmp(struct bucket *l, struct bucket *r)
> > {
> > - return l->gc_sectors_used < r->gc_sectors_used;
> > + return l->gc_sectors_dirty < r->gc_sectors_dirty;
> > }
> >
> > static unsigned int bucket_heap_top(struct cache *ca)
> > {
> > struct bucket *b;
> >
> > - return (b = heap_peek(&ca->heap)) ? b->gc_sectors_used : 0;
> > + return (b = heap_peek(&ca->heap)) ? b->gc_sectors_dirty : 0;
> > }
> >
> > void bch_moving_gc(struct cache_set *c)
> > @@ -215,17 +215,17 @@ void bch_moving_gc(struct cache_set *c)
> >
> > for_each_bucket(b, ca) {
> > if (GC_MOVE(b) || GC_MARK(b) == GC_MARK_METADATA ||
> > - !b->gc_sectors_used ||
> > - b->gc_sectors_used == ca->sb.bucket_size ||
> > + !b->gc_sectors_dirty ||
> > + b->gc_sectors_dirty == ca->sb.bucket_size ||
> > atomic_read(&b->pin))
> > continue;
> >
> > if (!heap_full(&ca->heap)) {
> > - sectors_to_move += b->gc_sectors_used;
> > + sectors_to_move += b->gc_sectors_dirty;
> > heap_add(&ca->heap, b, bucket_cmp);
> > } else if (bucket_cmp(b, heap_peek(&ca->heap))) {
> > sectors_to_move -= bucket_heap_top(ca);
> > - sectors_to_move += b->gc_sectors_used;
> > + sectors_to_move += b->gc_sectors_dirty;
> >
> > ca->heap.data[0] = b;
> > heap_sift(&ca->heap, 0, bucket_cmp);
> > @@ -237,7 +237,7 @@ void bch_moving_gc(struct cache_set *c)
> >
> > while (sectors_to_move > reserve_sectors) {
> > heap_pop(&ca->heap, b, bucket_cmp);
> > - sectors_to_move -= b->gc_sectors_used;
> > + sectors_to_move -= b->gc_sectors_dirty;
> > }
> >
> > while (heap_pop(&ca->heap, b, bucket_cmp))
> > --
> > 2.17.1.windows.2
> >
>
> --
> Coly Li
>
