Re: [f2fs-dev] [PATCH v1] f2fs: Improve large section GC by locating valid block segments

From: Daeho Jeong
Date: Mon Apr 14 2025 - 18:22:16 EST

Next message: Nathan Lynch: "Re: [PATCH] Revert "dmaengine: dmatest: Fix dmatest waiting less when interrupted""
Previous message: Randy Dunlap: "Re: linux-next: Tree for Apr 14 (drivers/mmc/host/sdhci-esdhc-imx.c)"
In reply to: Chao Yu: "Re: [f2fs-dev] [PATCH v1] f2fs: Improve large section GC by locating valid block segments"
Next in thread: yohan.joung: "Re: [f2fs-dev] [PATCH v1] f2fs: Improve large section GC by locating valid block segments"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]

On Fri, Apr 11, 2025 at 3:07 AM Chao Yu <chao@xxxxxxxxxx> wrote:
>
> Yohan,
>
> Sorry for the delay, will catch up this a little bit latter.
>
> On 2025/4/10 8:17, yohan.joung wrote:
> > hi jeageuk, chao
> > How about changing the large section gc in this direction?
> > Thanks
> >
> >> Change the large section GC to locate valid block segments instead of
> >> performing a sequential search. This modification enhances performance
> >> by reducing unnecessary block scanning in large storage sections.
> >>
> >> example :
> >> [invalid seg 0] [invalid seg 1] [invalid seg 2]
> >> [ valid seg 3] [ valid seg 4] [ valid seg 5]
> >>
> >> Current: In the first GC, nothing is moved,
> >> but in the second GC, segments 3, 4, and 5 are moved.
> >> Change: In the first GC, segments 3, 4, and 5 are moved.
> >>
> >> Signed-off-by: yohan.joung <yohan.joung@xxxxxx>
> >> ---
> >> fs/f2fs/f2fs.h | 2 ++
> >> fs/f2fs/gc.c | 92 +++++++++++++++++++++++++++++++++++--------------
> >> fs/f2fs/gc.h | 6 ++++
> >> fs/f2fs/super.c | 8 ++++-
> >> 4 files changed, 82 insertions(+), 26 deletions(-)
> >>
> >> diff --git a/fs/f2fs/f2fs.h b/fs/f2fs/f2fs.h
> >> index f1576dc6ec67..348417edac25 100644
> >> --- a/fs/f2fs/f2fs.h
> >> +++ b/fs/f2fs/f2fs.h
> >> @@ -4008,6 +4008,8 @@ int f2fs_gc_range(struct f2fs_sb_info *sbi,
> >> int f2fs_resize_fs(struct file *filp, __u64 block_count);
> >> int __init f2fs_create_garbage_collection_cache(void);
> >> void f2fs_destroy_garbage_collection_cache(void);
> >> +int __init f2fs_init_garbage_collection_summary_cache(void);
> >> +void f2fs_destroy_garbage_collection_summary_cache(void);
> >> /* victim selection function for cleaning and SSR */
> >> int f2fs_get_victim(struct f2fs_sb_info *sbi, unsigned int *result,
> >> int gc_type, int type, char alloc_mode,
> >> diff --git a/fs/f2fs/gc.c b/fs/f2fs/gc.c
> >> index 2b8f9239bede..3b63e85fa038 100644
> >> --- a/fs/f2fs/gc.c
> >> +++ b/fs/f2fs/gc.c
> >> @@ -24,6 +24,7 @@
> >> #include <trace/events/f2fs.h>
> >>
> >> static struct kmem_cache *victim_entry_slab;
> >> +static struct kmem_cache *gc_page_entry_slab;
> >>
> >> static unsigned int count_bits(const unsigned long *addr,
> >> unsigned int offset, unsigned int len);
> >> @@ -711,6 +712,30 @@ static void release_victim_entry(struct f2fs_sb_info *sbi)
> >> f2fs_bug_on(sbi, !list_empty(&am->victim_list));
> >> }
> >>
> >> +static struct gc_page_entry *add_gc_page_entry(struct list_head *gc_page_list,
> >> + struct page *sum_page, unsigned int segno)
> >> +{
> >> + struct gc_page_entry *gpe;
> >> +
> >> + gpe = f2fs_kmem_cache_alloc(gc_page_entry_slab, GFP_NOFS, true, NULL);
> >> + gpe->segno = segno;
> >> + gpe->sum_page = sum_page;
> >> + list_add_tail(&gpe->list, gc_page_list);
> >> + return gpe;
> >> +}
> >> +
> >> +static void release_gc_page_entry(struct list_head *gc_page_list, bool putpage)
> >> +{
> >> + struct gc_page_entry *gpe, *tmp;
> >> +
> >> + list_for_each_entry_safe(gpe, tmp, gc_page_list, list) {
> >> + if (putpage)
> >> + f2fs_put_page(gpe->sum_page, 0);
> >> + list_del(&gpe->list);
> >> + kmem_cache_free(gc_page_entry_slab, gpe);
> >> + }
> >> +}
> >> +
> >> static bool f2fs_pin_section(struct f2fs_sb_info *sbi, unsigned int segno)
> >> {
> >> struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
> >> @@ -1721,14 +1746,18 @@ static int do_garbage_collect(struct f2fs_sb_info *sbi,
> >> struct page *sum_page;
> >> struct f2fs_summary_block *sum;
> >> struct blk_plug plug;
> >> + struct gc_page_entry *gpe;
> >> unsigned int segno = start_segno;
> >> unsigned int end_segno = start_segno + SEGS_PER_SEC(sbi);
> >> unsigned int sec_end_segno;
> >> + unsigned int window_granularity = 1;
> >> int seg_freed = 0, migrated = 0;
> >> unsigned char type = IS_DATASEG(get_seg_entry(sbi, segno)->type) ?
> >> SUM_TYPE_DATA : SUM_TYPE_NODE;
> >> unsigned char data_type = (type == SUM_TYPE_DATA) ? DATA : NODE;
> >> int submitted = 0;
> >> + int gc_list_count = 0;
> >> + LIST_HEAD(gc_page_list);
> >>
> >> if (__is_large_section(sbi)) {
> >> sec_end_segno = rounddown(end_segno, SEGS_PER_SEC(sbi));
> >> @@ -1744,7 +1773,7 @@ static int do_garbage_collect(struct f2fs_sb_info *sbi,
> >> f2fs_usable_segs_in_sec(sbi);
> >>
> >> if (gc_type == BG_GC || one_time) {
> >> - unsigned int window_granularity =
> >> + window_granularity =
> >> sbi->migration_window_granularity;

Plz, refer to the below described in sysfs-fs-f2fs.

What: /sys/fs/f2fs/<disk>/migration_window_granularity
Date: September 2024
Contact: "Daeho Jeong" <daehojeong@xxxxxxxxxx>
Description: Controls migration window granularity of garbage
collection on large
section. it can control the scanning window
granularity for GC migration
in a unit of segment, while migration_granularity
controls the number
of segments which can be migrated at the same turn.

In your patch, it's more like migration_granularity.
Plus, I think we don't need migration_window_granularity anymore with
your patch.

> >>
> >> if (f2fs_sb_has_blkzoned(sbi) &&
> >> @@ -1752,8 +1781,6 @@ static int do_garbage_collect(struct f2fs_sb_info *sbi,
> >> sbi->gc_thread->boost_zoned_gc_percent))
> >> window_granularity *=
> >> BOOST_GC_MULTIPLE;
> >> -
> >> - end_segno = start_segno + window_granularity;
> >> }
> >>
> >> if (end_segno > sec_end_segno)
> >> @@ -1762,37 +1789,38 @@ static int do_garbage_collect(struct f2fs_sb_info *sbi,
> >>
> >> sanity_check_seg_type(sbi, get_seg_entry(sbi, segno)->type);
> >>
> >> - /* readahead multi ssa blocks those have contiguous address */
> >> - if (__is_large_section(sbi))
> >> + for (segno = start_segno; segno < end_segno; segno++) {
> >> + if (!get_valid_blocks(sbi, segno, false))
> >> + continue;
> >> +
> >> + /* readahead multi ssa blocks those have contiguous address */
> >> f2fs_ra_meta_pages(sbi, GET_SUM_BLOCK(sbi, segno),
> >> - end_segno - segno, META_SSA, true);
> >> + window_granularity, META_SSA, true);
> >>
> >> - /* reference all summary page */
> >> - while (segno < end_segno) {
> >> - sum_page = f2fs_get_sum_page(sbi, segno++);
> >> + /* reference all summary page */
> >> + sum_page = f2fs_get_sum_page(sbi, segno);
> >> if (IS_ERR(sum_page)) {
> >> int err = PTR_ERR(sum_page);
> >> -
> >> - end_segno = segno - 1;
> >> - for (segno = start_segno; segno < end_segno; segno++) {
> >> - sum_page = find_get_page(META_MAPPING(sbi),
> >> - GET_SUM_BLOCK(sbi, segno));
> >> - f2fs_put_page(sum_page, 0);
> >> - f2fs_put_page(sum_page, 0);
> >> - }
> >> + release_gc_page_entry(&gc_page_list, true);
> >> return err;
> >> }
> >> + add_gc_page_entry(&gc_page_list, sum_page, segno);
> >> unlock_page(sum_page);
> >> + if (++gc_list_count >= window_granularity)
> >> + break;
> >> }
> >>
> >> blk_start_plug(&plug);
> >>
> >> - for (segno = start_segno; segno < end_segno; segno++) {
> >> + list_for_each_entry(gpe, &gc_page_list, list) {
> >>
> >> /* find segment summary of victim */
> >> - sum_page = find_get_page(META_MAPPING(sbi),
> >> - GET_SUM_BLOCK(sbi, segno));
> >> - f2fs_put_page(sum_page, 0);
> >> + sum_page = gpe->sum_page;
> >> + segno = gpe->segno;
> >> + if (!sum_page) {
> >> + f2fs_err(sbi, "Failed to get summary page for segment %u", segno);
> >> + goto skip;
> >> + }
> >>
> >> if (get_valid_blocks(sbi, segno, false) == 0)
> >> goto freed;
> >> @@ -1835,18 +1863,20 @@ static int do_garbage_collect(struct f2fs_sb_info *sbi,
> >> get_valid_blocks(sbi, segno, false) == 0)
> >> seg_freed++;
> >>
> >> - if (__is_large_section(sbi))
> >> - sbi->next_victim_seg[gc_type] =
> >> - (segno + 1 < sec_end_segno) ?
> >> - segno + 1 : NULL_SEGNO;
> >> skip:
> >> f2fs_put_page(sum_page, 0);
> >> }
> >>
> >> + if (__is_large_section(sbi) && !list_empty(&gc_page_list))
> >> + sbi->next_victim_seg[gc_type] =
> >> + (segno + 1 < sec_end_segno) ?
> >> + segno + 1 : NULL_SEGNO;
> >> +
> >> if (submitted)
> >> f2fs_submit_merged_write(sbi, data_type);
> >>
> >> blk_finish_plug(&plug);
> >> + release_gc_page_entry(&gc_page_list, false);
> >>
> >> if (migrated)
> >> stat_inc_gc_sec_count(sbi, data_type, gc_type);
> >> @@ -2008,6 +2038,18 @@ int f2fs_gc(struct f2fs_sb_info *sbi, struct f2fs_gc_control *gc_control)
> >> return ret;
> >> }
> >>
> >> +int __init f2fs_init_garbage_collection_summary_cache(void)
> >> +{
> >> + gc_page_entry_slab = f2fs_kmem_cache_create("f2fs_gc_page_entry",
> >> + sizeof(struct gc_page_entry));
> >> + return gc_page_entry_slab ? 0 : -ENOMEM;
> >> +}
> >> +
> >> +void f2fs_destroy_garbage_collection_summary_cache(void)
> >> +{
> >> + kmem_cache_destroy(gc_page_entry_slab);
> >> +}
> >> +
> >> int __init f2fs_create_garbage_collection_cache(void)
> >> {
> >> victim_entry_slab = f2fs_kmem_cache_create("f2fs_victim_entry",
> >> diff --git a/fs/f2fs/gc.h b/fs/f2fs/gc.h
> >> index 5c1eaf55e127..9c8695efe394 100644
> >> --- a/fs/f2fs/gc.h
> >> +++ b/fs/f2fs/gc.h
> >> @@ -82,6 +82,12 @@ struct victim_entry {
> >> struct list_head list;
> >> };
> >>
> >> +struct gc_page_entry {
> >> + unsigned int segno;
> >> + struct page *sum_page;
> >> + struct list_head list;
> >> +};
> >> +
> >> /*
> >> * inline functions
> >> */
> >> diff --git a/fs/f2fs/super.c b/fs/f2fs/super.c
> >> index f087b2b71c89..a3241730fe4f 100644
> >> --- a/fs/f2fs/super.c
> >> +++ b/fs/f2fs/super.c
> >> @@ -5090,9 +5090,12 @@ static int __init init_f2fs_fs(void)
> >> err = f2fs_create_garbage_collection_cache();
> >> if (err)
> >> goto free_extent_cache;
> >> - err = f2fs_init_sysfs();
> >> + err = f2fs_init_garbage_collection_summary_cache();
> >> if (err)
> >> goto free_garbage_collection_cache;
> >> + err = f2fs_init_sysfs();
> >> + if (err)
> >> + goto free_garbage_collection_summary_cache;
> >> err = f2fs_init_shrinker();
> >> if (err)
> >> goto free_sysfs;
> >> @@ -5141,6 +5144,8 @@ static int __init init_f2fs_fs(void)
> >> f2fs_exit_shrinker();
> >> free_sysfs:
> >> f2fs_exit_sysfs();
> >> +free_garbage_collection_summary_cache:
> >> + f2fs_destroy_garbage_collection_summary_cache();
> >> free_garbage_collection_cache:
> >> f2fs_destroy_garbage_collection_cache();
> >> free_extent_cache:
> >> @@ -5172,6 +5177,7 @@ static void __exit exit_f2fs_fs(void)
> >> f2fs_destroy_root_stats();
> >> f2fs_exit_shrinker();
> >> f2fs_exit_sysfs();
> >> + f2fs_destroy_garbage_collection_summary_cache();
> >> f2fs_destroy_garbage_collection_cache();
> >> f2fs_destroy_extent_cache();
> >> f2fs_destroy_recovery_cache();
> >> --
> >> 2.33.0
> >>

Do you have any numbers showing how much the performance is enhanced?

Thanks.

> >>
> >>
> >> _______________________________________________
> >> Linux-f2fs-devel mailing list
> >> Linux-f2fs-devel@xxxxxxxxxxxxxxxxxxxxx
>

Next message: Nathan Lynch: "Re: [PATCH] Revert "dmaengine: dmatest: Fix dmatest waiting less when interrupted""
Previous message: Randy Dunlap: "Re: linux-next: Tree for Apr 14 (drivers/mmc/host/sdhci-esdhc-imx.c)"
In reply to: Chao Yu: "Re: [f2fs-dev] [PATCH v1] f2fs: Improve large section GC by locating valid block segments"
Next in thread: yohan.joung: "Re: [f2fs-dev] [PATCH v1] f2fs: Improve large section GC by locating valid block segments"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]