Message-ID: <0bb733a8-5f86-950f-9474-edde718b1546@kernel.org>
Date:   Fri, 18 Aug 2017 22:18:50 +0800
From:   Chao Yu <chao@...nel.org>
To:     Jaegeuk Kim <jaegeuk@...nel.org>, Chao Yu <yuchao0@...wei.com>
Cc:     linux-f2fs-devel@...ts.sourceforge.net,
        linux-kernel@...r.kernel.org
Subject: Re: [PATCH v4] f2fs: introduce discard_granularity sysfs entry

Hi Jaegeuk,

Sorry for the delay, the modification looks good to me. ;)

Thanks,

On 2017/8/16 1:54, Jaegeuk Kim wrote:
> On 08/15, Chao Yu wrote:
>> On 2017/8/15 11:45, Jaegeuk Kim wrote:
>>> On 08/07, Chao Yu wrote:
>>>> From: Chao Yu <yuchao0@...wei.com>
>>>>
>>>> Commit d618ebaf0aa8 ("f2fs: enable small discard by default") enables
>>>> f2fs to issue 4K size discard in real-time discard mode. However, issuing
>>>> smaller discard may cost more lifetime but releasing less free space in
>>>> flash device. Since f2fs has ability of separating hot/cold data and
>>>> garbage collection, we can expect that small-sized invalid region would
>>>> expand soon with OPU, deletion or garbage collection on valid datas, so
>>>> it's better to delay or skip issuing smaller size discards, it could help
>>>> to reduce overmuch consumption of IO bandwidth and lifetime of flash
>>>> storage.
>>>>
>>>> This patch makes f2fs selectng 64K size as its default minimal
>>>> granularity, and issue discard with the size which is not smaller than
>>>> minimal granularity. Also it exposes discard granularity as sysfs entry
>>>> for configuration in different scenario.
>>>
>>> Hi Chao,
>>>
>>> I'd like to change the default value to 1 in order to keep the original
>>> behavior, since we must avoid performance fluctuation after this single
>>> patch. Instead, you probably can change the value through sysfs.
>>
>> As I know, in fragmented filesystem space, there are may dozens of thousand
>> discard, in scenario of cellphone user are using, 30% is above 64K size, but
>> occupy 75% space of all undiscard space, so I changed discard_granularity to 64K
>> just to release bulk space in device. For other small-sized discards, I expect
>> that they may extend and cross the granularity threshold soon, and fstrim of
>> android could cover them in the night.
> 
> Yup, I thought that, but this patch prevents fstrim from issuing small discards
> due to the granularity check. And, low-end device likes to issue small discards
> much more. How about this?
> 
> From a0f38a8574a35995ba9e9e81ae5138919bb672a8 Mon Sep 17 00:00:00 2001
> From: Chao Yu <yuchao0@...wei.com>
> Date: Mon, 7 Aug 2017 23:09:56 +0800
> Subject: [PATCH] f2fs: introduce discard_granularity sysfs entry
> 
> Commit d618ebaf0aa8 ("f2fs: enable small discard by default") enables
> f2fs to issue 4K size discard in real-time discard mode. However, issuing
> smaller discard may cost more lifetime but releasing less free space in
> flash device. Since f2fs has ability of separating hot/cold data and
> garbage collection, we can expect that small-sized invalid region would
> expand soon with OPU, deletion or garbage collection on valid datas, so
> it's better to delay or skip issuing smaller size discards, it could help
> to reduce overmuch consumption of IO bandwidth and lifetime of flash
> storage.
> 
> This patch makes f2fs selectng 64K size as its default minimal
> granularity, and issue discard with the size which is not smaller than
> minimal granularity. Also it exposes discard granularity as sysfs entry
> for configuration in different scenario.
> 
> Jaegeuk Kim:
>  We must issue all the accumulated discard commands when fstrim is called.
>  So, I've added pend_list_tag[] to indicate whether we should issue the
>  commands or not. If tag sets P_ACTIVE or P_TRIM, we have to issue them.
>  P_TRIM is set once at a time, given fstrim trigger.
> 
> Signed-off-by: Chao Yu <yuchao0@...wei.com>
> Signed-off-by: Jaegeuk Kim <jaegeuk@...nel.org>
> ---
>  Documentation/ABI/testing/sysfs-fs-f2fs |  9 +++++++
>  fs/f2fs/f2fs.h                          |  9 +++++++
>  fs/f2fs/segment.c                       | 43 +++++++++++++++++++++++++++++++--
>  fs/f2fs/sysfs.c                         | 23 ++++++++++++++++++
>  4 files changed, 82 insertions(+), 2 deletions(-)
> 
> diff --git a/Documentation/ABI/testing/sysfs-fs-f2fs b/Documentation/ABI/testing/sysfs-fs-f2fs
> index 621da3fc56c5..11b7f4ebea7c 100644
> --- a/Documentation/ABI/testing/sysfs-fs-f2fs
> +++ b/Documentation/ABI/testing/sysfs-fs-f2fs
> @@ -57,6 +57,15 @@ Contact:	"Jaegeuk Kim" <jaegeuk.kim@...sung.com>
>  Description:
>  		 Controls the issue rate of small discard commands.
>  
> +What:          /sys/fs/f2fs/<disk>/discard_granularity
> +Date:          July 2017
> +Contact:       "Chao Yu" <yuchao0@...wei.com>
> +Description:
> +		Controls discard granularity of inner discard thread, inner thread
> +		will not issue discards with size that is smaller than granularity.
> +		The unit size is one block, now only support configuring in range
> +		of [1, 512].
> +
>  What:		/sys/fs/f2fs/<disk>/max_victim_search
>  Date:		January 2014
>  Contact:	"Jaegeuk Kim" <jaegeuk.kim@...sung.com>
> diff --git a/fs/f2fs/f2fs.h b/fs/f2fs/f2fs.h
> index e252e5bf9791..336021b9b93e 100644
> --- a/fs/f2fs/f2fs.h
> +++ b/fs/f2fs/f2fs.h
> @@ -196,11 +196,18 @@ struct discard_entry {
>  	unsigned char discard_map[SIT_VBLOCK_MAP_SIZE];	/* segment discard bitmap */
>  };
>  
> +/* default discard granularity of inner discard thread, unit: block count */
> +#define DEFAULT_DISCARD_GRANULARITY		16
> +
>  /* max discard pend list number */
>  #define MAX_PLIST_NUM		512
>  #define plist_idx(blk_num)	((blk_num) >= MAX_PLIST_NUM ?		\
>  					(MAX_PLIST_NUM - 1) : (blk_num - 1))
>  
> +#define P_ACTIVE	0x01
> +#define P_TRIM		0x02
> +#define plist_issue(tag)	(((tag) & P_ACTIVE) || ((tag) & P_TRIM))
> +
>  enum {
>  	D_PREP,
>  	D_SUBMIT,
> @@ -236,11 +243,13 @@ struct discard_cmd_control {
>  	struct task_struct *f2fs_issue_discard;	/* discard thread */
>  	struct list_head entry_list;		/* 4KB discard entry list */
>  	struct list_head pend_list[MAX_PLIST_NUM];/* store pending entries */
> +	unsigned char pend_list_tag[MAX_PLIST_NUM];/* tag for pending entries */
>  	struct list_head wait_list;		/* store on-flushing entries */
>  	wait_queue_head_t discard_wait_queue;	/* waiting queue for wake-up */
>  	struct mutex cmd_lock;
>  	unsigned int nr_discards;		/* # of discards in the list */
>  	unsigned int max_discards;		/* max. discards to be issued */
> +	unsigned int discard_granularity;	/* discard granularity */
>  	unsigned int undiscard_blks;		/* # of undiscard blocks */
>  	atomic_t issued_discard;		/* # of issued discard */
>  	atomic_t issing_discard;		/* # of issing discard */
> diff --git a/fs/f2fs/segment.c b/fs/f2fs/segment.c
> index 05144b3a7f62..8c90b69dcd6d 100644
> --- a/fs/f2fs/segment.c
> +++ b/fs/f2fs/segment.c
> @@ -1028,22 +1028,49 @@ static void __issue_discard_cmd(struct f2fs_sb_info *sbi, bool issue_cond)
>  	f2fs_bug_on(sbi,
>  		!__check_rb_tree_consistence(sbi, &dcc->root));
>  	blk_start_plug(&plug);
> -	for (i = MAX_PLIST_NUM - 1; i >= 0; i--) {
> +	for (i = MAX_PLIST_NUM - 1;
> +			i >= 0 && plist_issue(dcc->pend_list_tag[i]); i--) {
>  		pend_list = &dcc->pend_list[i];
>  		list_for_each_entry_safe(dc, tmp, pend_list, list) {
>  			f2fs_bug_on(sbi, dc->state != D_PREP);
>  
> +			/* Hurry up to finish fstrim */
> +			if (dcc->pend_list_tag[i] & P_TRIM) {
> +				__submit_discard_cmd(sbi, dc);
> +				continue;
> +			}
> +
>  			if (!issue_cond || is_idle(sbi))
>  				__submit_discard_cmd(sbi, dc);
>  			if (issue_cond && iter++ > DISCARD_ISSUE_RATE)
>  				goto out;
>  		}
> +		if (list_empty(pend_list) && dcc->pend_list_tag[i] & P_TRIM)
> +			dcc->pend_list_tag[i] &= (~P_TRIM);
>  	}
>  out:
>  	blk_finish_plug(&plug);
>  	mutex_unlock(&dcc->cmd_lock);
>  }
>  
> +static void __drop_discard_cmd(struct f2fs_sb_info *sbi)
> +{
> +	struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info;
> +	struct list_head *pend_list;
> +	struct discard_cmd *dc, *tmp;
> +	int i;
> +
> +	mutex_lock(&dcc->cmd_lock);
> +	for (i = MAX_PLIST_NUM - 1; i >= 0; i--) {
> +		pend_list = &dcc->pend_list[i];
> +		list_for_each_entry_safe(dc, tmp, pend_list, list) {
> +			f2fs_bug_on(sbi, dc->state != D_PREP);
> +			__remove_discard_cmd(sbi, dc);
> +		}
> +	}
> +	mutex_unlock(&dcc->cmd_lock);
> +}
> +
>  static void __wait_one_discard_bio(struct f2fs_sb_info *sbi,
>  							struct discard_cmd *dc)
>  {
> @@ -1126,6 +1153,7 @@ void stop_discard_thread(struct f2fs_sb_info *sbi)
>  void f2fs_wait_discard_bios(struct f2fs_sb_info *sbi)
>  {
>  	__issue_discard_cmd(sbi, false);
> +	__drop_discard_cmd(sbi);
>  	__wait_discard_cmd(sbi, false);
>  }
>  
> @@ -1448,9 +1476,13 @@ static int create_discard_cmd_control(struct f2fs_sb_info *sbi)
>  	if (!dcc)
>  		return -ENOMEM;
>  
> +	dcc->discard_granularity = DEFAULT_DISCARD_GRANULARITY;
>  	INIT_LIST_HEAD(&dcc->entry_list);
> -	for (i = 0; i < MAX_PLIST_NUM; i++)
> +	for (i = 0; i < MAX_PLIST_NUM; i++) {
>  		INIT_LIST_HEAD(&dcc->pend_list[i]);
> +		if (i >= dcc->discard_granularity - 1)
> +			dcc->pend_list_tag[i] |= P_ACTIVE;
> +	}
>  	INIT_LIST_HEAD(&dcc->wait_list);
>  	mutex_init(&dcc->cmd_lock);
>  	atomic_set(&dcc->issued_discard, 0);
> @@ -2079,11 +2111,13 @@ bool exist_trim_candidates(struct f2fs_sb_info *sbi, struct cp_control *cpc)
>  
>  int f2fs_trim_fs(struct f2fs_sb_info *sbi, struct fstrim_range *range)
>  {
> +	struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info;
>  	__u64 start = F2FS_BYTES_TO_BLK(range->start);
>  	__u64 end = start + F2FS_BYTES_TO_BLK(range->len) - 1;
>  	unsigned int start_segno, end_segno;
>  	struct cp_control cpc;
>  	int err = 0;
> +	int i;
>  
>  	if (start >= MAX_BLKADDR(sbi) || range->len < sbi->blocksize)
>  		return -EINVAL;
> @@ -2127,6 +2161,11 @@ int f2fs_trim_fs(struct f2fs_sb_info *sbi, struct fstrim_range *range)
>  
>  		schedule();
>  	}
> +	/* It's time to issue all the filed discards */
> +	mutex_lock(&dcc->cmd_lock);
> +	for (i = 0; i < MAX_PLIST_NUM; i++)
> +		dcc->pend_list_tag[i] |= P_TRIM;
> +	mutex_unlock(&dcc->cmd_lock);
>  out:
>  	range->len = F2FS_BLK_TO_BYTES(cpc.trimmed);
>  	return err;
> diff --git a/fs/f2fs/sysfs.c b/fs/f2fs/sysfs.c
> index c40e5d24df9f..4bcaa9059026 100644
> --- a/fs/f2fs/sysfs.c
> +++ b/fs/f2fs/sysfs.c
> @@ -152,6 +152,27 @@ static ssize_t f2fs_sbi_store(struct f2fs_attr *a,
>  		spin_unlock(&sbi->stat_lock);
>  		return count;
>  	}
> +
> +	if (!strcmp(a->attr.name, "discard_granularity")) {
> +		struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info;
> +		int i;
> +
> +		if (t == 0 || t > MAX_PLIST_NUM)
> +			return -EINVAL;
> +		if (t == *ui)
> +			return count;
> +
> +		mutex_lock(&dcc->cmd_lock);
> +		for (i = 0; i < MAX_PLIST_NUM; i++) {
> +			if (i >= t - 1)
> +				dcc->pend_list_tag[i] |= P_ACTIVE;
> +			else
> +				dcc->pend_list_tag[i] &= (~P_ACTIVE);
> +		}
> +		mutex_unlock(&dcc->cmd_lock);
> +		return count;
> +	}
> +
>  	*ui = t;
>  
>  	if (!strcmp(a->attr.name, "iostat_enable") && *ui == 0)
> @@ -248,6 +269,7 @@ F2FS_RW_ATTR(GC_THREAD, f2fs_gc_kthread, gc_idle, gc_idle);
>  F2FS_RW_ATTR(GC_THREAD, f2fs_gc_kthread, gc_urgent, gc_urgent);
>  F2FS_RW_ATTR(SM_INFO, f2fs_sm_info, reclaim_segments, rec_prefree_segments);
>  F2FS_RW_ATTR(DCC_INFO, discard_cmd_control, max_small_discards, max_discards);
> +F2FS_RW_ATTR(DCC_INFO, discard_cmd_control, discard_granularity, discard_granularity);
>  F2FS_RW_ATTR(RESERVED_BLOCKS, f2fs_sb_info, reserved_blocks, reserved_blocks);
>  F2FS_RW_ATTR(SM_INFO, f2fs_sm_info, batched_trim_sections, trim_sections);
>  F2FS_RW_ATTR(SM_INFO, f2fs_sm_info, ipu_policy, ipu_policy);
> @@ -290,6 +312,7 @@ static struct attribute *f2fs_attrs[] = {
>  	ATTR_LIST(gc_urgent),
>  	ATTR_LIST(reclaim_segments),
>  	ATTR_LIST(max_small_discards),
> +	ATTR_LIST(discard_granularity),
>  	ATTR_LIST(batched_trim_sections),
>  	ATTR_LIST(ipu_policy),
>  	ATTR_LIST(min_ipu_util),
> 

Powered by blists - more mailing lists