RE:(2) [PATCH v4] f2fs: add async reset zone command support
From: Daejun Park
Date: Wed May 03 2023 - 22:09:23 EST
Thank you for reviewing the patch.
> --------- Original Message ---------
> Sender : Chao Yu <chao@xxxxxxxxxx>
> Date : 2023-05-04 10:43 (GMT+9)
> Title : Re: [PATCH v4] f2fs: add async reset zone command support
> To : 박대준<daejun7.park@xxxxxxxxxxx>, jaegeuk@xxxxxxxxxx<jaegeuk@xxxxxxxxxx>, rostedt@xxxxxxxxxxx<rostedt@xxxxxxxxxxx>, mhiramat@xxxxxxxxxx<mhiramat@xxxxxxxxxx>, linux-f2fs-devel@xxxxxxxxxxxxxxxxxxxxx<linux-f2fs-devel@xxxxxxxxxxxxxxxxxxxxx>, linux-kernel@xxxxxxxxxxxxxxx<linux-kernel@xxxxxxxxxxxxxxx>, linux-trace-kernel@xxxxxxxxxxxxxxx<linux-trace-kernel@xxxxxxxxxxxxxxx>
> CC : 김석환<sukka.kim@xxxxxxxxxxx>, 송용길<yonggil.song@xxxxxxxxxxx>, 김범수<beomsu7.kim@xxxxxxxxxxx>
>
> On 2023/5/2 12:16, Daejun Park wrote:
> > Changelog:
> >
> > v3 -> v4
> > Fixed build error caused by unused function.
> >
> > v2 -> v3
> > Modified arguments to be correct for ftrace parameter.
> > Changed __submit_zone_reset_cmd to void return.
> > Refactored the f2fs_wait_discard_bio function.
> > Fixed code that was previously incorrectly merged.
> >
> > v1 -> v2
> > Changed to apply the optional async reset write pointer by default.
> >
> > 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>
> > ---
> > fs/f2fs/segment.c | 83 +++++++++++++++++++++++++++++++++++--
> > include/trace/events/f2fs.h | 18 +++++++-
> > 2 files changed, 96 insertions(+), 5 deletions(-)
> >
> > diff --git a/fs/f2fs/segment.c b/fs/f2fs/segment.c
> > index 6db410f1bb8c..ec7a8de71198 100644
> > --- a/fs/f2fs/segment.c
> > +++ b/fs/f2fs/segment.c
> > @@ -1196,6 +1196,44 @@ 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 void __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, 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);
>
> How about accounting iostat info for zone reset command?
>
> f2fs_update_iostat(sbi, NULL, FS_ZONE_RESET_IO, len * F2FS_BLKSIZE);
Sure, I will do.
> > +
> > + atomic_inc(&dcc->issued_discard);
> > +}
> > +#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,
> > @@ -1217,6 +1255,13 @@ 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)) {
> > + __submit_zone_reset_cmd(sbi, dc, flag, wait_list, issued);
> > + return 0;
> > + }
> > +#endif
> > +
> > trace_f2fs_issue_discard(bdev, dc->di.start, dc->di.len);
> >
> > lstart = dc->di.lstart;
> > @@ -1461,6 +1506,19 @@ 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 lblkstart,
> > + block_t blklen)
> > +{
> > + 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)
> > {
> > @@ -1724,6 +1782,19 @@ static void f2fs_wait_discard_bio(struct f2fs_sb_info *sbi, block_t blkaddr)
> >
> > mutex_lock(&dcc->cmd_lock);
> > dc = __lookup_discard_cmd(sbi, blkaddr);
> > +#ifdef CONFIG_BLK_DEV_ZONED
> > + if (dc && f2fs_sb_has_blkzoned(sbi) && bdev_is_zoned(dc->bdev)) {
> > + /* force submit zone reset */
> > + if (dc->state == D_PREP)
> > + __submit_zone_reset_cmd(sbi, dc, REQ_SYNC,
> > + &dcc->wait_list, NULL);
> > + dc->ref++;
>
> IIUC, should be?
>
> else
> dc->ref++;
>
> Thanks,
In case of discard, we can delete a pending discard via __punch_discard_cmd().
However, we cannot remove the pending zone reset command.
And, we should always wait for the completion of the zone reset command,
so I increased the reference counter.
Thanks,
Daejun
> > + mutex_unlock(&dcc->cmd_lock);
> > + /* wait zone reset */
> > + __wait_one_discard_bio(sbi, dc);
> > + return;
> > + }
> > +#endif
> > if (dc) {
> > if (dc->state == D_PREP) {
> > __punch_discard_cmd(sbi, dc, blkaddr);
> > @@ -1876,9 +1947,15 @@ static int __f2fs_issue_discard_zone(struct f2fs_sb_info *sbi,
> > blkstart, blklen);
> > return -EIO;
> > }
> > - trace_f2fs_issue_reset_zone(bdev, blkstart);
> > - return blkdev_zone_mgmt(bdev, REQ_OP_ZONE_RESET,
> > - sector, nr_sects, GFP_NOFS);
> > +
> > + if (unlikely(is_sbi_flag_set(sbi, SBI_POR_DOING))) {
> > + trace_f2fs_issue_reset_zone(bdev, blkstart);
> > + return blkdev_zone_mgmt(bdev, REQ_OP_ZONE_RESET,
> > + sector, nr_sects, GFP_NOFS);
> > + }
> > +
> > + __queue_zone_reset_cmd(sbi, bdev, blkstart, lblkstart, blklen);
> > + return 0;
> > }
> >
> > /* For conventional zones, use regular discard if supported */
> > 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,
>
>
>