[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <ee0b8b83-a368-4529-b25d-d3b1c6291648@kernel.org>
Date: Thu, 7 Aug 2025 16:38:25 +0800
From: Chao Yu <chao@...nel.org>
To: Liao Yuanhong <liaoyuanhong@...o.com>, Jaegeuk Kim <jaegeuk@...nel.org>
Cc: chao@...nel.org, linux-f2fs-devel@...ts.sourceforge.net,
linux-kernel@...r.kernel.org
Subject: Re: [PATCH v2] f2fs: Add bggc_block_io to adjust the priority of
BG_GC when issuing IO
On 8/6/25 15:09, Liao Yuanhong wrote:
> Currently, we have encountered some issues while testing ZUFS. In
> situations near the storage limit (e.g., 50GB remaining), and after
> simulating fragmentation by repeatedly writing and deleting data, we found
> that application installation and startup tests conducted after idling for
> a few minutes take significantly longer several times that of traditional
> UFS. Tracing the operations revealed that the majority of I/Os were issued
> by background GC, which blocks normal I/O operations.
>
> Under normal circumstances, ZUFS indeed requires more background GC and
> employs a more aggressive GC strategy. However, I aim to find a way to
> minimize the impact on regular I/O operations under these near-limit
> conditions. To address this, I have introduced a bggc_block_io feature,
> which controls the prioritization of background GC in the presence of I/Os.
> This switch can be adjusted at the framework level to implement different
> strategies. If set to ALL_IO_PRIOR, all background GC operations will be
> skipped during active I/O issuance. The default option remains consistent
> with the current strategy, ensuring no change in behavior.
>
> Signed-off-by: Liao Yuanhong <liaoyuanhong@...o.com>
> ---
> Changes in v2:
> Non ZUFS can also be adjusted through this option.
> ---
> Documentation/ABI/testing/sysfs-fs-f2fs | 13 +++++++++++++
> fs/f2fs/f2fs.h | 18 +++++++++++-------
> fs/f2fs/super.c | 2 ++
> fs/f2fs/sysfs.c | 9 +++++++++
> 4 files changed, 35 insertions(+), 7 deletions(-)
>
> diff --git a/Documentation/ABI/testing/sysfs-fs-f2fs b/Documentation/ABI/testing/sysfs-fs-f2fs
> index bc0e7fefc39d..12fda11d4da5 100644
> --- a/Documentation/ABI/testing/sysfs-fs-f2fs
> +++ b/Documentation/ABI/testing/sysfs-fs-f2fs
> @@ -883,3 +883,16 @@ Date: June 2025
> Contact: "Daeho Jeong" <daehojeong@...gle.com>
> Description: Control GC algorithm for boost GC. 0: cost benefit, 1: greedy
> Default: 1
> +
> +What: /sys/fs/f2fs/<disk>/bggc_block_io
> +Date: August 2025
> +Contact: "Liao Yuanhong" <liaoyuanhong@...o.com>
> +Description: Used to adjust the BG_GC priority when issuing IO, with a default value
> + of 1.
Default value is 2 if CONFIG_BLK_DEV_ZONED is off?
Thanks,
> +
> + ================== =============================================
> + value description
> + bggc_block_io = 0 Prioritize background GC
> + bggc_block_io = 1 Stop background GC only when issuing read I/O
> + bggc_block_io = 2 Stop background GC when issuing I/O
> + ================== =============================================
> diff --git a/fs/f2fs/f2fs.h b/fs/f2fs/f2fs.h
> index 46be7560548c..fe41b733fbc2 100644
> --- a/fs/f2fs/f2fs.h
> +++ b/fs/f2fs/f2fs.h
> @@ -155,6 +155,12 @@ enum blkzone_allocation_policy {
> BLKZONE_ALLOC_PRIOR_CONV, /* Prioritize writing to conventional zones */
> };
>
> +enum bggc_block_io_policy {
> + BGGC_PRIOR,
> + READ_IO_PRIOR,
> + ALL_IO_PRIOR,
> +};
> +
> /*
> * An implementation of an rwsem that is explicitly unfair to readers. This
> * prevents priority inversion when a low-priority reader acquires the read lock
> @@ -1804,6 +1810,7 @@ struct f2fs_sb_info {
> spinlock_t dev_lock; /* protect dirty_device */
> bool aligned_blksize; /* all devices has the same logical blksize */
> unsigned int first_seq_zone_segno; /* first segno in sequential zone */
> + unsigned int bggc_block_io; /* For adjust the BG_GC priority when issuing IO */
>
> /* For write statistics */
> u64 sectors_written_start;
> @@ -2998,13 +3005,10 @@ static inline bool is_idle(struct f2fs_sb_info *sbi, int type)
> if (sbi->gc_mode == GC_URGENT_HIGH)
> return true;
>
> - if (zoned_gc) {
> - if (is_inflight_read_io(sbi))
> - return false;
> - } else {
> - if (is_inflight_io(sbi, type))
> - return false;
> - }
> + if (sbi->bggc_block_io == READ_IO_PRIOR && is_inflight_read_io(sbi))
> + return false;
> + if (sbi->bggc_block_io == ALL_IO_PRIOR && is_inflight_io(sbi, type))
> + return false;
>
> if (sbi->gc_mode == GC_URGENT_MID)
> return true;
> diff --git a/fs/f2fs/super.c b/fs/f2fs/super.c
> index e16c4e2830c2..a21cecc5424e 100644
> --- a/fs/f2fs/super.c
> +++ b/fs/f2fs/super.c
> @@ -4629,9 +4629,11 @@ static int f2fs_scan_devices(struct f2fs_sb_info *sbi)
>
> logical_blksize = bdev_logical_block_size(sbi->sb->s_bdev);
> sbi->aligned_blksize = true;
> + sbi->bggc_block_io = ALL_IO_PRIOR;
> #ifdef CONFIG_BLK_DEV_ZONED
> sbi->max_open_zones = UINT_MAX;
> sbi->blkzone_alloc_policy = BLKZONE_ALLOC_PRIOR_SEQ;
> + sbi->bggc_block_io = READ_IO_PRIOR;
> #endif
>
> for (i = 0; i < max_devices; i++) {
> diff --git a/fs/f2fs/sysfs.c b/fs/f2fs/sysfs.c
> index f736052dea50..efea15209788 100644
> --- a/fs/f2fs/sysfs.c
> +++ b/fs/f2fs/sysfs.c
> @@ -674,6 +674,13 @@ static ssize_t __sbi_store(struct f2fs_attr *a,
> sbi->blkzone_alloc_policy = t;
> return count;
> }
> +
> + if (!strcmp(a->attr.name, "bggc_block_io")) {
> + if (t < BGGC_PRIOR || t > ALL_IO_PRIOR)
> + return -EINVAL;
> + sbi->bggc_block_io = t;
> + return count;
> + }
> #endif
>
> #ifdef CONFIG_F2FS_FS_COMPRESSION
> @@ -1172,6 +1179,7 @@ F2FS_SBI_GENERAL_RW_ATTR(max_read_extent_count);
> #ifdef CONFIG_BLK_DEV_ZONED
> F2FS_SBI_GENERAL_RO_ATTR(unusable_blocks_per_sec);
> F2FS_SBI_GENERAL_RW_ATTR(blkzone_alloc_policy);
> +F2FS_SBI_GENERAL_RW_ATTR(bggc_block_io);
> #endif
> F2FS_SBI_GENERAL_RW_ATTR(carve_out);
> F2FS_SBI_GENERAL_RW_ATTR(reserved_pin_section);
> @@ -1342,6 +1350,7 @@ static struct attribute *f2fs_attrs[] = {
> #ifdef CONFIG_BLK_DEV_ZONED
> ATTR_LIST(unusable_blocks_per_sec),
> ATTR_LIST(blkzone_alloc_policy),
> + ATTR_LIST(bggc_block_io),
> #endif
> #ifdef CONFIG_F2FS_FS_COMPRESSION
> ATTR_LIST(compr_written_block),
Powered by blists - more mailing lists