Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Date:   Tue, 14 Aug 2018 10:23:13 -0700
From:   Jaegeuk Kim <jaegeuk@kernel.org>
To:     Chao Yu <yuchao0@huawei.com>
Cc:     linux-f2fs-devel@lists.sourceforge.net,
        linux-kernel@vger.kernel.org, chao@kernel.org
Subject: Re: [PATCH 2/2] f2fs: tune discard speed with storage usage rate
Message-ID: <20180814172313.GC56510@jaegeuk-macbookpro.roam.corp.google.com>
References: <20180810100806.9298-1-yuchao0@huawei.com>
 <20180810100806.9298-2-yuchao0@huawei.com>
 <20180814041906.GC52730@jaegeuk-macbookpro.roam.corp.google.com>
 <57d9b6ea-68a5-4736-0b34-74db539d8959@huawei.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <57d9b6ea-68a5-4736-0b34-74db539d8959@huawei.com>
User-Agent: Mutt/1.8.2 (2017-04-18)
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On 08/14, Chao Yu wrote:
> On 2018/8/14 12:19, Jaegeuk Kim wrote:
> > On 08/10, Chao Yu wrote:
> >> Previously, discard speed was fixed mostly, and in high usage rate
> >> device, we will speed up issuing discard, but it doesn't make sense
> >> that in a non-full filesystem, we still issue discard with slow speed.
> > 
> > Could you please elaborate the problem in more detail? The speed depends
> > on how many candidates?
> 
> undiscard blocks are all 4k granularity.
> a) utility: filesystem: 20% + undiscard blocks: 20% = flash storage: 40%
> b) utility: filesystem: 40% + undiscard blocks: 25% = flash storage: 65%
> c) utility: filesystem: 60% + undiscard blocks: 30% = flash storage: 100%
> 
> 
> 1. for case c), we need to speed up issuing discard based on utilization of
> "filesystem + undiscard" instead of just utilization of filesystem.
> 
> -		if (utilization(sbi) > DEF_DISCARD_URGENT_UTIL) {
> -			dpolicy->granularity = 1;
> -			dpolicy->max_interval = DEF_MIN_DISCARD_ISSUE_TIME;
> -		}
> 
> 2. If free space in storage touches therein threshold, performance will be very
> sensitive. In low-end storage, with high usage in space, even free space is
> reduced by 1%, performance will decrease a lot.

So, we may need to distinguish low-end vs. high-end storage. In high-end case,
it'd be better to avoid IO contention, while low-end device wants to get more
discard commands as much as possible. So, how about adding an option for this
as a tunable point?

> 
> IMO, in above cases, we'd better to issue discard with high speed for c), middle
> speed for b), and low speed for a).
> 
> How do you think?
> 
> Thanks,
> 
> > 
> > Thanks,
> > 
> >>
> >> Anyway, it comes out undiscarded block makes FTL GC be lower efficient
> >> and causing high lifetime overhead.
> >>
> >> Let's tune discard speed as below:
> >>
> >> a. adjust default issue interval:
> >> 		original	after
> >> min_interval:	50ms		100ms
> >> mid_interval:	500ms		1000ms
> >> max_interval:	60000ms		10000ms
> >>
> >> b. if last time we stop issuing discard due to IO interruption of user,
> >> let's reset all {min,mid,max}_interval to default one.
> >>
> >> c. tune {min,mid,max}_interval with below calculation method:
> >>
> >> base_interval = default_interval / 10;
> >> total_interval = default_interval - base_interval;
> >> interval = base_interval + total_interval * (100 - dev_util) / 100;
> >>
> >> For example:
> >> min_interval (:100ms)
> >> dev_util (%)	interval (ms)
> >> 0		100
> >> 10		91
> >> 20		82
> >> 30		73
> >> ...
> >> 80		28
> >> 90		19
> >> 100		10
> >>
> >> Signed-off-by: Chao Yu <yuchao0@huawei.com>
> >> ---
> >>  fs/f2fs/f2fs.h    | 11 ++++----
> >>  fs/f2fs/segment.c | 64 +++++++++++++++++++++++++++++++++++++----------
> >>  fs/f2fs/segment.h |  9 +++++++
> >>  fs/f2fs/super.c   |  2 +-
> >>  4 files changed, 67 insertions(+), 19 deletions(-)
> >>
> >> diff --git a/fs/f2fs/f2fs.h b/fs/f2fs/f2fs.h
> >> index 273ffdaf4891..a1dd2e1c3cb9 100644
> >> --- a/fs/f2fs/f2fs.h
> >> +++ b/fs/f2fs/f2fs.h
> >> @@ -185,10 +185,9 @@ enum {
> >>  
> >>  #define MAX_DISCARD_BLOCKS(sbi)		BLKS_PER_SEC(sbi)
> >>  #define DEF_MAX_DISCARD_REQUEST		8	/* issue 8 discards per round */
> >> -#define DEF_MIN_DISCARD_ISSUE_TIME	50	/* 50 ms, if exists */
> >> -#define DEF_MID_DISCARD_ISSUE_TIME	500	/* 500 ms, if device busy */
> >> -#define DEF_MAX_DISCARD_ISSUE_TIME	60000	/* 60 s, if no candidates */
> >> -#define DEF_DISCARD_URGENT_UTIL		80	/* do more discard over 80% */
> >> +#define DEF_MIN_DISCARD_ISSUE_TIME	100	/* 100 ms, if exists */
> >> +#define DEF_MID_DISCARD_ISSUE_TIME	1000	/* 1000 ms, if device busy */
> >> +#define DEF_MAX_DISCARD_ISSUE_TIME	10000	/* 10000 ms, if no candidates */
> >>  #define DEF_CP_INTERVAL			60	/* 60 secs */
> >>  #define DEF_IDLE_INTERVAL		5	/* 5 secs */
> >>  
> >> @@ -248,7 +247,8 @@ struct discard_entry {
> >>  };
> >>  
> >>  /* default discard granularity of inner discard thread, unit: block count */
> >> -#define DEFAULT_DISCARD_GRANULARITY		1
> >> +#define MID_DISCARD_GRANULARITY			16
> >> +#define MIN_DISCARD_GRANULARITY			1
> >>  
> >>  /* max discard pend list number */
> >>  #define MAX_PLIST_NUM		512
> >> @@ -330,6 +330,7 @@ struct discard_cmd_control {
> >>  	atomic_t discard_cmd_cnt;		/* # of cached cmd count */
> >>  	struct rb_root root;			/* root of discard rb-tree */
> >>  	bool rbtree_check;			/* config for consistence check */
> >> +	bool io_interrupted;			/* last state of io interrupted */
> >>  };
> >>  
> >>  /* for the list of fsync inodes, used only during recovery */
> >> diff --git a/fs/f2fs/segment.c b/fs/f2fs/segment.c
> >> index 8b52e8dfb12f..9564aaf1f27b 100644
> >> --- a/fs/f2fs/segment.c
> >> +++ b/fs/f2fs/segment.c
> >> @@ -968,6 +968,44 @@ static void __check_sit_bitmap(struct f2fs_sb_info *sbi,
> >>  #endif
> >>  }
> >>  
> >> +static void __adjust_discard_speed(unsigned int *interval,
> >> +				unsigned int def_interval, int dev_util)
> >> +{
> >> +	unsigned int base_interval, total_interval;
> >> +
> >> +	base_interval = def_interval / 10;
> >> +	total_interval = def_interval - base_interval;
> >> +
> >> +	/*
> >> +	 * if def_interval = 100, adjusted interval should be in range of
> >> +	 * [10, 100].
> >> +	 */
> >> +	*interval = base_interval + total_interval * (100 - dev_util) / 100;
> >> +}
> >> +
> >> +static void __tune_discard_policy(struct f2fs_sb_info *sbi,
> >> +					struct discard_policy *dpolicy)
> >> +{
> >> +	struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info;
> >> +	int dev_util;
> >> +
> >> +	if (dcc->io_interrupted) {
> >> +		dpolicy->min_interval = DEF_MIN_DISCARD_ISSUE_TIME;
> >> +		dpolicy->mid_interval = DEF_MID_DISCARD_ISSUE_TIME;
> >> +		dpolicy->max_interval = DEF_MAX_DISCARD_ISSUE_TIME;
> >> +		return;
> >> +	}
> >> +
> >> +	dev_util = dev_utilization(sbi);
> >> +
> >> +	__adjust_discard_speed(&dpolicy->min_interval,
> >> +				DEF_MIN_DISCARD_ISSUE_TIME, dev_util);
> >> +	__adjust_discard_speed(&dpolicy->mid_interval,
> >> +				DEF_MID_DISCARD_ISSUE_TIME, dev_util);
> >> +	__adjust_discard_speed(&dpolicy->max_interval,
> >> +				DEF_MAX_DISCARD_ISSUE_TIME, dev_util);
> >> +}
> >> +
> >>  static void __init_discard_policy(struct f2fs_sb_info *sbi,
> >>  				struct discard_policy *dpolicy,
> >>  				int discard_type, unsigned int granularity)
> >> @@ -982,20 +1020,11 @@ static void __init_discard_policy(struct f2fs_sb_info *sbi,
> >>  	dpolicy->io_aware_gran = MAX_PLIST_NUM;
> >>  
> >>  	if (discard_type == DPOLICY_BG) {
> >> -		dpolicy->min_interval = DEF_MIN_DISCARD_ISSUE_TIME;
> >> -		dpolicy->mid_interval = DEF_MID_DISCARD_ISSUE_TIME;
> >> -		dpolicy->max_interval = DEF_MAX_DISCARD_ISSUE_TIME;
> >>  		dpolicy->io_aware = true;
> >>  		dpolicy->sync = false;
> >>  		dpolicy->ordered = true;
> >> -		if (utilization(sbi) > DEF_DISCARD_URGENT_UTIL) {
> >> -			dpolicy->granularity = 1;
> >> -			dpolicy->max_interval = DEF_MIN_DISCARD_ISSUE_TIME;
> >> -		}
> >> +		__tune_discard_policy(sbi, dpolicy);
> >>  	} else if (discard_type == DPOLICY_FORCE) {
> >> -		dpolicy->min_interval = DEF_MIN_DISCARD_ISSUE_TIME;
> >> -		dpolicy->mid_interval = DEF_MID_DISCARD_ISSUE_TIME;
> >> -		dpolicy->max_interval = DEF_MAX_DISCARD_ISSUE_TIME;
> >>  		dpolicy->io_aware = false;
> >>  	} else if (discard_type == DPOLICY_FSTRIM) {
> >>  		dpolicy->io_aware = false;
> >> @@ -1353,6 +1382,8 @@ static unsigned int __issue_discard_cmd_orderly(struct f2fs_sb_info *sbi,
> >>  	if (!issued && io_interrupted)
> >>  		issued = -1;
> >>  
> >> +	dcc->io_interrupted = io_interrupted;
> >> +
> >>  	return issued;
> >>  }
> >>  
> >> @@ -1370,7 +1401,7 @@ static int __issue_discard_cmd(struct f2fs_sb_info *sbi,
> >>  		if (i + 1 < dpolicy->granularity)
> >>  			break;
> >>  
> >> -		if (i < DEFAULT_DISCARD_GRANULARITY && dpolicy->ordered)
> >> +		if (i < MID_DISCARD_GRANULARITY && dpolicy->ordered)
> >>  			return __issue_discard_cmd_orderly(sbi, dpolicy);
> >>  
> >>  		pend_list = &dcc->pend_list[i];
> >> @@ -1407,6 +1438,8 @@ static int __issue_discard_cmd(struct f2fs_sb_info *sbi,
> >>  	if (!issued && io_interrupted)
> >>  		issued = -1;
> >>  
> >> +	dcc->io_interrupted = io_interrupted;
> >> +
> >>  	return issued;
> >>  }
> >>  
> >> @@ -1576,7 +1609,11 @@ static int issue_discard_thread(void *data)
> >>  	struct f2fs_sb_info *sbi = data;
> >>  	struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info;
> >>  	wait_queue_head_t *q = &dcc->discard_wait_queue;
> >> -	struct discard_policy dpolicy;
> >> +	struct discard_policy dpolicy = {
> >> +		.min_interval = DEF_MIN_DISCARD_ISSUE_TIME,
> >> +		.mid_interval = DEF_MID_DISCARD_ISSUE_TIME,
> >> +		.max_interval = DEF_MAX_DISCARD_ISSUE_TIME,
> >> +	};
> >>  	unsigned int wait_ms = DEF_MIN_DISCARD_ISSUE_TIME;
> >>  	int issued;
> >>  
> >> @@ -1929,7 +1966,7 @@ static int create_discard_cmd_control(struct f2fs_sb_info *sbi)
> >>  	if (!dcc)
> >>  		return -ENOMEM;
> >>  
> >> -	dcc->discard_granularity = DEFAULT_DISCARD_GRANULARITY;
> >> +	dcc->discard_granularity = MIN_DISCARD_GRANULARITY;
> >>  	INIT_LIST_HEAD(&dcc->entry_list);
> >>  	for (i = 0; i < MAX_PLIST_NUM; i++)
> >>  		INIT_LIST_HEAD(&dcc->pend_list[i]);
> >> @@ -1945,6 +1982,7 @@ static int create_discard_cmd_control(struct f2fs_sb_info *sbi)
> >>  	dcc->next_pos = 0;
> >>  	dcc->root = RB_ROOT;
> >>  	dcc->rbtree_check = false;
> >> +	dcc->io_interrupted = false;
> >>  
> >>  	init_waitqueue_head(&dcc->discard_wait_queue);
> >>  	SM_I(sbi)->dcc_info = dcc;
> >> diff --git a/fs/f2fs/segment.h b/fs/f2fs/segment.h
> >> index 422b0ceb1eaa..63b4da72cd34 100644
> >> --- a/fs/f2fs/segment.h
> >> +++ b/fs/f2fs/segment.h
> >> @@ -616,6 +616,15 @@ static inline int utilization(struct f2fs_sb_info *sbi)
> >>  					sbi->user_block_count);
> >>  }
> >>  
> >> +static inline int dev_utilization(struct f2fs_sb_info *sbi)
> >> +{
> >> +	unsigned int dev_blks;
> >> +
> >> +	dev_blks = valid_user_blocks(sbi) + SM_I(sbi)->dcc_info->undiscard_blks;
> >> +	return div_u64((u64)dev_blks * 100,
> >> +			MAIN_SEGS(sbi) << sbi->log_blocks_per_seg);
> >> +}
> >> +
> >>  /*
> >>   * Sometimes f2fs may be better to drop out-of-place update policy.
> >>   * And, users can control the policy through sysfs entries.
> >> diff --git a/fs/f2fs/super.c b/fs/f2fs/super.c
> >> index b055f2ea77c5..55ed76daad23 100644
> >> --- a/fs/f2fs/super.c
> >> +++ b/fs/f2fs/super.c
> >> @@ -2862,7 +2862,7 @@ static void f2fs_tuning_parameters(struct f2fs_sb_info *sbi)
> >>  	/* adjust parameters according to the volume size */
> >>  	if (sm_i->main_segments <= SMALL_VOLUME_SEGMENTS) {
> >>  		F2FS_OPTION(sbi).alloc_mode = ALLOC_MODE_REUSE;
> >> -		sm_i->dcc_info->discard_granularity = 1;
> >> +		sm_i->dcc_info->discard_granularity = MIN_DISCARD_GRANULARITY;
> >>  		sm_i->ipu_policy = 1 << F2FS_IPU_FORCE;
> >>  	}
> >>  
> >> -- 
> >> 2.18.0.rc1
> > 
> > .
> >