RE:(2) [PATCH] f2fs: add async reset zone command support

From: Daejun Park
Date: Thu Apr 20 2023 - 03:55:13 EST


Hi Jeageuk,

Thanks for helpful review.

>
> On 04/14, Daejun Park wrote:
> > This patch enables submit reset zone command asynchornously. It helps
> > decrease average latency of write IOs in high utilization scenario by
> > faster checkpointing.
> >
> > Signed-off-by: Daejun Park <daejun7.park@xxxxxxxxxxx>
> > ---
> >  Documentation/filesystems/f2fs.rst |  4 ++
> >  fs/f2fs/f2fs.h                     |  1 +
> >  fs/f2fs/segment.c                  | 92 +++++++++++++++++++++++++++++-
> >  fs/f2fs/super.c                    |  8 +++
> >  include/trace/events/f2fs.h        | 18 +++++-
> >  5 files changed, 119 insertions(+), 4 deletions(-)
> >
> > diff --git a/Documentation/filesystems/f2fs.rst b/Documentation/filesystems/f2fs.rst
> > index 2055e72871fe..4cfabf831a79 100644
> > --- a/Documentation/filesystems/f2fs.rst
> > +++ b/Documentation/filesystems/f2fs.rst
> > @@ -342,6 +342,10 @@ discard_unit=%s                 Control discard unit, the argument can be "block", "segment"
> >                           default, it is helpful for large sized SMR or ZNS devices to
> >                           reduce memory cost by getting rid of fs metadata supports small
> >                           discard.
> > +async_reset_zone         Enable the RESET WRITE POINTER command to be submitted asynchronously.
> > +                         With this option, the RESET WRITE POINTER command can be processed by
> > +                         the discard thread like a discard command. It can reduce checkpoint
> > +                         latency by asynchronously checking for completion of the RESET WRITE POINTER command.
>
> Do we need to set this by default instead of mount option?

I think it is good idea. I will make this feature by default.

>
> >  memory=%s                 Control memory mode. This supports "normal" and "low" modes.
> >                           "low" mode is introduced to support low memory devices.
> >                           Because of the nature of low memory devices, in this mode, f2fs
> > diff --git a/fs/f2fs/f2fs.h b/fs/f2fs/f2fs.h
> > index 4e2596dacbf1..021e55c5d1a8 100644
> > --- a/fs/f2fs/f2fs.h
> > +++ b/fs/f2fs/f2fs.h
> > @@ -167,6 +167,7 @@ struct f2fs_mount_info {
> >                                           * be aligned to this unit: block,
> >                                           * segment or section
> >                                           */
> > +        bool async_zone_reset;          /* async zone reset */
> >          struct fscrypt_dummy_policy dummy_enc_policy; /* test dummy encryption */
> >          block_t unusable_cap_perc;        /* percentage for cap */
> >          block_t unusable_cap;                /* Amount of space allowed to be
> > diff --git a/fs/f2fs/segment.c b/fs/f2fs/segment.c
> > index 45128694eefa..60cfe97c9dbd 100644
> > --- a/fs/f2fs/segment.c
> > +++ b/fs/f2fs/segment.c
> > @@ -1189,6 +1189,46 @@ static void __init_discard_policy(struct f2fs_sb_info *sbi,
> >  static void __update_discard_tree_range(struct f2fs_sb_info *sbi,
> >                                  struct block_device *bdev, block_t lstart,
> >                                  block_t start, block_t len);
> > +
> > +#ifdef CONFIG_BLK_DEV_ZONED
> > +static int __submit_zone_reset_cmd(struct f2fs_sb_info *sbi,
> > +                                   struct discard_cmd *dc, blk_opf_t flag,
> > +                                   struct list_head *wait_list,
> > +                                   unsigned int *issued)
> > +{
> > +        struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info;
> > +        struct block_device *bdev = dc->bdev;
> > +        struct bio *bio = bio_alloc(bdev, 0, REQ_OP_ZONE_RESET | flag, GFP_NOFS);
> > +        unsigned long flags;
> > +
> > +        trace_f2fs_issue_reset_zone(bdev, SECTOR_FROM_BLOCK(dc->di.start));
> > +
> > +        spin_lock_irqsave(&dc->lock, flags);
> > +        dc->state = D_SUBMIT;
> > +        dc->bio_ref++;
> > +        spin_unlock_irqrestore(&dc->lock, flags);
> > +
> > +        if (issued)
> > +                (*issued)++;
> > +
> > +        atomic_inc(&dcc->queued_discard);
> > +        dc->queued++;
> > +        list_move_tail(&dc->list, wait_list);
> > +
> > +        /* sanity check on discard range */
> > +        __check_sit_bitmap(sbi, dc->di.lstart, dc->di.lstart + dc->di.len);
> > +
> > +        bio->bi_iter.bi_sector = SECTOR_FROM_BLOCK(dc->di.start);
> > +        bio->bi_private = dc;
> > +        bio->bi_end_io = f2fs_submit_discard_endio;
> > +        submit_bio(bio);
> > +
> > +        atomic_inc(&dcc->issued_discard);
> > +
> > +        return 0;
> > +}
> > +#endif
> > +
> >  /* this function is copied from blkdev_issue_discard from block/blk-lib.c */
> >  static int __submit_discard_cmd(struct f2fs_sb_info *sbi,
> >                                  struct discard_policy *dpolicy,
> > @@ -1210,6 +1250,11 @@ static int __submit_discard_cmd(struct f2fs_sb_info *sbi,
> >          if (is_sbi_flag_set(sbi, SBI_NEED_FSCK))
> >                  return 0;
> >  
> > +#ifdef CONFIG_BLK_DEV_ZONED
> > +        if (f2fs_sb_has_blkzoned(sbi) && bdev_is_zoned(bdev))
> > +                return __submit_zone_reset_cmd(sbi, dc, flag, wait_list, issued);
> > +#endif
> > +
> >          trace_f2fs_issue_discard(bdev, dc->di.start, dc->di.len);
> >  
> >          lstart = dc->di.lstart;
> > @@ -1454,21 +1499,42 @@ static void __update_discard_tree_range(struct f2fs_sb_info *sbi,
> >          }
> >  }
> >  
> > +#ifdef CONFIG_BLK_DEV_ZONED
> > +static void __queue_zone_reset_cmd(struct f2fs_sb_info *sbi,
> > +                struct block_device *bdev, block_t blkstart, block_t blklen)
> > +{
> > +        block_t lblkstart = blkstart;
> > +
> > +        if (f2fs_is_multi_device(sbi)) {
> > +                int devi = f2fs_target_device_index(sbi, blkstart);
> > +
> > +                blkstart -= FDEV(devi).start_blk;
> > +        }
> > +
> > +        trace_f2fs_queue_reset_zone(bdev, blkstart);
> > +
> > +        mutex_lock(&SM_I(sbi)->dcc_info->cmd_lock);
> > +        __insert_discard_cmd(sbi, bdev, lblkstart, blkstart, blklen);
> > +        mutex_unlock(&SM_I(sbi)->dcc_info->cmd_lock);
> > +}
> > +#endif
> > +
> >  static void __queue_discard_cmd(struct f2fs_sb_info *sbi,
> >                  struct block_device *bdev, block_t blkstart, block_t blklen)
> >  {
> >          block_t lblkstart = blkstart;
> >  
> > +        trace_f2fs_queue_discard(bdev, blkstart, blklen);
> > +
> >          if (!f2fs_bdev_support_discard(bdev))
> >                  return;
> >  
> > -        trace_f2fs_queue_discard(bdev, blkstart, blklen);
> > -
> >          if (f2fs_is_multi_device(sbi)) {
> >                  int devi = f2fs_target_device_index(sbi, blkstart);
> >  
> >                  blkstart -= FDEV(devi).start_blk;
> >          }
> > +
> >          mutex_lock(&SM_I(sbi)->dcc_info->cmd_lock);
> >          __update_discard_tree_range(sbi, bdev, lblkstart, blkstart, blklen);
> >          mutex_unlock(&SM_I(sbi)->dcc_info->cmd_lock);
> > @@ -1719,7 +1785,22 @@ static void f2fs_wait_discard_bio(struct f2fs_sb_info *sbi, block_t blkaddr)
> >          dc = __lookup_discard_cmd(sbi, blkaddr);
> >          if (dc) {
> >                  if (dc->state == D_PREP) {
> > +#ifdef CONFIG_BLK_DEV_ZONED
> > +                        if (f2fs_sb_has_blkzoned(sbi) &&
> > +                            bdev_is_zoned(dc->bdev)) {
> > +                                struct discard_policy dpolicy;
> > +                                /* force submit zone reset */
> > +                                __init_discard_policy(sbi, &dpolicy,
> > +                                                      DPOLICY_FORCE, 1);
> > +                                __submit_discard_cmd(sbi, &dpolicy, dc, NULL);
> > +                                dc->ref++;
> > +                                need_wait = true;
> > +                        } else {
> > +                                __punch_discard_cmd(sbi, dc, blkaddr);
>
> Can be consolidated in both cases below.

How do you think this modification as below.

if (dc) {
if (dc->state == D_PREP) {
if (f2fs_sb_has_blkzoned(sbi) &&
bdev_is_zoned(dc->bdev)) {
struct discard_policy dpolicy;
/* force submit zone reset */
__init_discard_policy(sbi, &dpolicy,
DPOLICY_FORCE, 1);
__submit_discard_cmd(sbi, &dpolicy, dc, NULL);
dc->ref++;
need_wait = true;
} else {
__punch_discard_cmd(sbi, dc, blkaddr);
}
} else {
dc->ref++;
need_wait = true;
}
}

>
> > +                        }
> > +#else
> >                          __punch_discard_cmd(sbi, dc, blkaddr);
> > +#endif
> >                  } else {
> >                          dc->ref++;
> >                          need_wait = true;
> > @@ -1869,6 +1950,13 @@ static int __f2fs_issue_discard_zone(struct f2fs_sb_info *sbi,
> >                                   blkstart, blklen);
> >                          return -EIO;
> >                  }
> > +
> > +                if (likely(!is_sbi_flag_set(sbi, SBI_POR_DOING)) &&
> > +                           F2FS_OPTION(sbi).async_zone_reset) {
> > +                        __queue_zone_reset_cmd(sbi, bdev, lblkstart, blklen);
> > +                        return 0;
> > +                }
>
> Can make a sigle case above and below, if we set this by default.

I can remove "F2FS_OPTION(sbi).async_zone_reset", but I think the checking
"SBI_POR_DOING" statement can not be removed. Because check_zone_write_pointer()
and fix_curseg_write_pointer() use this __f2fs_issue_discard_zone() for sending
reset write pointer command synchronously.

> > +
> >                  trace_f2fs_issue_reset_zone(bdev, blkstart);
> >                  return blkdev_zone_mgmt(bdev, REQ_OP_ZONE_RESET,
> >                                          sector, nr_sects, GFP_NOFS);
> > diff --git a/fs/f2fs/super.c b/fs/f2fs/super.c
> > index 7d0202f7b317..48198112cbbc 100644
> > --- a/fs/f2fs/super.c
> > +++ b/fs/f2fs/super.c
> > @@ -162,6 +162,7 @@ enum {
> >          Opt_gc_merge,
> >          Opt_nogc_merge,
> >          Opt_discard_unit,
> > +        Opt_async_zone_reset,
> >          Opt_memory_mode,
> >          Opt_age_extent_cache,
> >          Opt_err,
> > @@ -241,6 +242,7 @@ static match_table_t f2fs_tokens = {
> >          {Opt_gc_merge, "gc_merge"},
> >          {Opt_nogc_merge, "nogc_merge"},
> >          {Opt_discard_unit, "discard_unit=%s"},
> > +        {Opt_async_zone_reset, "async_zone_reset"},
> >          {Opt_memory_mode, "memory=%s"},
> >          {Opt_age_extent_cache, "age_extent_cache"},
> >          {Opt_err, NULL},
> > @@ -1249,6 +1251,9 @@ static int parse_options(struct super_block *sb, char *options, bool is_remount)
> >                          }
> >                          kfree(name);
> >                          break;
> > +                case Opt_async_zone_reset:
> > +                        F2FS_OPTION(sbi).async_zone_reset = true;
> > +                        break;
> >                  case Opt_memory_mode:
> >                          name = match_strdup(&args[0]);
> >                          if (!name)
> > @@ -2047,6 +2052,9 @@ static int f2fs_show_options(struct seq_file *seq, struct dentry *root)
> >          if (test_opt(sbi, ATGC))
> >                  seq_puts(seq, ",atgc");
> >  
> > +        if (F2FS_OPTION(sbi).async_zone_reset)
> > +                seq_puts(seq, ",async_zone_reset");
> > +
> >          if (F2FS_OPTION(sbi).memory_mode == MEMORY_MODE_NORMAL)
> >                  seq_printf(seq, ",memory=%s", "normal");
> >          else if (F2FS_OPTION(sbi).memory_mode == MEMORY_MODE_LOW)
> > diff --git a/include/trace/events/f2fs.h b/include/trace/events/f2fs.h
> > index 99cbc5949e3c..ee1477de8324 100644
> > --- a/include/trace/events/f2fs.h
> > +++ b/include/trace/events/f2fs.h
> > @@ -1512,7 +1512,7 @@ DEFINE_EVENT(f2fs_discard, f2fs_remove_discard,
> >          TP_ARGS(dev, blkstart, blklen)
> >  );
> >  
> > -TRACE_EVENT(f2fs_issue_reset_zone,
> > +DECLARE_EVENT_CLASS(f2fs_reset_zone,
> >  
> >          TP_PROTO(struct block_device *dev, block_t blkstart),
> >  
> > @@ -1528,11 +1528,25 @@ TRACE_EVENT(f2fs_issue_reset_zone,
> >                  __entry->blkstart = blkstart;
> >          ),
> >  
> > -        TP_printk("dev = (%d,%d), reset zone at block = 0x%llx",
> > +        TP_printk("dev = (%d,%d), zone at block = 0x%llx",
> >                  show_dev(__entry->dev),
> >                  (unsigned long long)__entry->blkstart)
> >  );
> >  
> > +DEFINE_EVENT(f2fs_reset_zone, f2fs_queue_reset_zone,
> > +
> > +        TP_PROTO(struct block_device *dev, block_t blkstart),
> > +
> > +        TP_ARGS(dev, blkstart)
> > +);
> > +
> > +DEFINE_EVENT(f2fs_reset_zone, f2fs_issue_reset_zone,
> > +
> > +        TP_PROTO(struct block_device *dev, block_t blkstart),
> > +
> > +        TP_ARGS(dev, blkstart)
> > +);
> > +
> >  TRACE_EVENT(f2fs_issue_flush,
> >  
> >          TP_PROTO(struct block_device *dev, unsigned int nobarrier,
> > --
> > 2.25.1
>

Thanks,
Daejun