Date:   Thu, 27 Sep 2018 10:42:38 +0530
From:   Sahitya Tummala <stummala@...eaurora.org>
To:     Chao Yu <chao@...nel.org>
Cc:     jaegeuk@...nel.org, linux-kernel@...r.kernel.org,
        linux-f2fs-devel@...ts.sourceforge.net
Subject: Re: [f2fs-dev] [PATCH v4] f2fs: allow out-place-update for direct IO
 in LFS mode

On Fri, Sep 21, 2018 at 09:12:22PM +0800, Chao Yu wrote:
> From: Chao Yu <yuchao0@...wei.com>
> 
> Normally, DIO uses in-pllace-update, but in LFS mode, f2fs doesn't
> allow triggering any in-place-update writes, so we fallback direct
> write to buffered write, result in bad performance of large size
> write.
> 
> This patch adds to support triggering out-place-update for direct IO
> to enhance its performance.
> 
> Note that it needs to exclude direct read IO during direct write,
> since new data writing to new block address will no be valid until
> write finished.
> 
> storage: zram
> 
> time xfs_io -f -d /mnt/f2fs/file -c "pwrite 0 1073741824" -c "fsync"
> 
> Before:
> real	0m13.061s
> user	0m0.327s
> sys	0m12.486s
> 
> After:
> real	0m6.448s
> user	0m0.228s
> sys	0m6.212s
> 
> Signed-off-by: Chao Yu <yuchao0@...wei.com>
> ---
> v4:
> - correct parameter in f2fs_sb_has_blkzoned()
>  fs/f2fs/data.c | 44 +++++++++++++++++++++++++++++++++++---------
>  fs/f2fs/f2fs.h | 45 +++++++++++++++++++++++++++++++++++++++++----
>  fs/f2fs/file.c |  3 ++-
>  3 files changed, 78 insertions(+), 14 deletions(-)
> 
> diff --git a/fs/f2fs/data.c b/fs/f2fs/data.c
> index b96f8588d565..38d5baa1c35d 100644
> --- a/fs/f2fs/data.c
> +++ b/fs/f2fs/data.c
> @@ -894,7 +894,7 @@ static int __allocate_data_block(struct dnode_of_data *dn, int seg_type)
>  
>  	dn->data_blkaddr = datablock_addr(dn->inode,
>  				dn->node_page, dn->ofs_in_node);
> -	if (dn->data_blkaddr == NEW_ADDR)
> +	if (dn->data_blkaddr != NULL_ADDR)
>  		goto alloc;
>  
>  	if (unlikely((err = inc_valid_block_count(sbi, dn->inode, &count))))
> @@ -950,7 +950,7 @@ int f2fs_preallocate_blocks(struct kiocb *iocb, struct iov_iter *from)
>  
>  	if (direct_io) {
>  		map.m_seg_type = f2fs_rw_hint_to_seg_type(iocb->ki_hint);
> -		flag = f2fs_force_buffered_io(inode, WRITE) ?
> +		flag = f2fs_force_buffered_io(inode, iocb, from) ?
>  					F2FS_GET_BLOCK_PRE_AIO :
>  					F2FS_GET_BLOCK_PRE_DIO;
>  		goto map_blocks;
> @@ -1069,7 +1069,15 @@ int f2fs_map_blocks(struct inode *inode, struct f2fs_map_blocks *map,
>  		goto sync_out;
>  	}
>  
> -	if (!is_valid_data_blkaddr(sbi, blkaddr)) {
> +	if (is_valid_data_blkaddr(sbi, blkaddr)) {
> +		/* use out-place-update for driect IO under LFS mode */
> +		if (test_opt(sbi, LFS) && create &&
> +				flag == F2FS_GET_BLOCK_DEFAULT) {

One of the recent patches from Jaegeuk, 0a4daae5ffea ("f2fs: update i_size after DIO
completion") added new flag for DIO - F2FS_GET_BLOCK_DIO. I think this patch needs to be
updated accordingly.

> +			err = __allocate_data_block(&dn, map->m_seg_type);
> +			if (!err)
> +				set_inode_flag(inode, FI_APPEND_WRITE);
> +		}
> +	} else {
>  		if (create) {
>  			if (unlikely(f2fs_cp_error(sbi))) {
>  				err = -EIO;
> @@ -2493,36 +2501,53 @@ static ssize_t f2fs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
>  	struct address_space *mapping = iocb->ki_filp->f_mapping;
>  	struct inode *inode = mapping->host;
>  	struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
> +	struct f2fs_inode_info *fi = F2FS_I(inode);
>  	size_t count = iov_iter_count(iter);
>  	loff_t offset = iocb->ki_pos;
>  	int rw = iov_iter_rw(iter);
>  	int err;
>  	enum rw_hint hint = iocb->ki_hint;
>  	int whint_mode = F2FS_OPTION(sbi).whint_mode;
> +	bool do_opu;
>  
>  	err = check_direct_IO(inode, iter, offset);
>  	if (err)
>  		return err < 0 ? err : 0;
>  
> -	if (f2fs_force_buffered_io(inode, rw))
> +	if (f2fs_force_buffered_io(inode, iocb, iter))
>  		return 0;
>  
> +	do_opu = allow_outplace_dio(inode, iocb, iter);
> +
>  	trace_f2fs_direct_IO_enter(inode, offset, count, rw);
>  
>  	if (rw == WRITE && whint_mode == WHINT_MODE_OFF)
>  		iocb->ki_hint = WRITE_LIFE_NOT_SET;
>  
> -	if (!down_read_trylock(&F2FS_I(inode)->i_gc_rwsem[rw])) {
> -		if (iocb->ki_flags & IOCB_NOWAIT) {
> +	if (iocb->ki_flags & IOCB_NOWAIT) {
> +		if (!down_read_trylock(&fi->i_gc_rwsem[rw])) {
> +			iocb->ki_hint = hint;
> +			err = -EAGAIN;
> +			goto out;
> +		}
> +		if (do_opu && !down_read_trylock(&fi->i_gc_rwsem[READ])) {
> +			up_read(&fi->i_gc_rwsem[rw]);
>  			iocb->ki_hint = hint;
>  			err = -EAGAIN;
>  			goto out;
>  		}
> -		down_read(&F2FS_I(inode)->i_gc_rwsem[rw]);
> +	} else {
> +		down_read(&fi->i_gc_rwsem[rw]);
> +		if (do_opu)
> +			down_read(&fi->i_gc_rwsem[READ]);
>  	}
>  
>  	err = blockdev_direct_IO(iocb, inode, iter, get_data_block_dio);
> -	up_read(&F2FS_I(inode)->i_gc_rwsem[rw]);
> +
> +	if (do_opu)
> +		up_read(&fi->i_gc_rwsem[READ]);
> +
> +	up_read(&fi->i_gc_rwsem[rw]);
>  
>  	if (rw == WRITE) {
>  		if (whint_mode == WHINT_MODE_OFF)
> @@ -2530,7 +2555,8 @@ static ssize_t f2fs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
>  		if (err > 0) {
>  			f2fs_update_iostat(F2FS_I_SB(inode), APP_DIRECT_IO,
>  									err);
> -			set_inode_flag(inode, FI_UPDATE_WRITE);
> +			if (!do_opu)
> +				set_inode_flag(inode, FI_UPDATE_WRITE);
>  		} else if (err < 0) {
>  			f2fs_write_failed(mapping, offset + count);
>  		}
> diff --git a/fs/f2fs/f2fs.h b/fs/f2fs/f2fs.h
> index 894a2503e722..72d46860cee3 100644
> --- a/fs/f2fs/f2fs.h
> +++ b/fs/f2fs/f2fs.h
> @@ -8,6 +8,7 @@
>  #ifndef _LINUX_F2FS_H
>  #define _LINUX_F2FS_H
>  
> +#include <linux/uio.h>
>  #include <linux/types.h>
>  #include <linux/page-flags.h>
>  #include <linux/buffer_head.h>
> @@ -3486,11 +3487,47 @@ static inline bool f2fs_may_encrypt(struct inode *inode)
>  #endif
>  }
>  
> -static inline bool f2fs_force_buffered_io(struct inode *inode, int rw)
> +static inline int block_unaligned_IO(struct inode *inode,
> +				struct kiocb *iocb, struct iov_iter *iter)
>  {
> -	return (f2fs_post_read_required(inode) ||
> -			(rw == WRITE && test_opt(F2FS_I_SB(inode), LFS)) ||
> -			F2FS_I_SB(inode)->s_ndevs);
> +	unsigned int i_blkbits = READ_ONCE(inode->i_blkbits);
> +	unsigned int blocksize_mask = (1 << i_blkbits) - 1;
> +	loff_t offset = iocb->ki_pos;
> +	unsigned long align = offset | iov_iter_alignment(iter);
> +
> +	return align & blocksize_mask;
> +}
> +
> +static inline int allow_outplace_dio(struct inode *inode,
> +				struct kiocb *iocb, struct iov_iter *iter)
> +{
> +	struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
> +	int rw = iov_iter_rw(iter);
> +
> +	return (test_opt(sbi, LFS) && (rw == WRITE) &&
> +				!block_unaligned_IO(inode, iocb, iter));
> +}
> +
> +static inline bool f2fs_force_buffered_io(struct inode *inode,
> +				struct kiocb *iocb, struct iov_iter *iter)
> +{
> +	struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
> +	int rw = iov_iter_rw(iter);
> +
> +	if (f2fs_post_read_required(inode))
> +		return true;
> +	if (sbi->s_ndevs)
> +		return true;
> +	/*
> +	 * for blkzoned device, fallback direct IO to buffered IO, so
> +	 * all IOs can be serialized by log-structured write.
> +	 */
> +	if (f2fs_sb_has_blkzoned(sbi->sb))
> +		return true;
> +	if (test_opt(sbi, LFS) && (rw == WRITE) &&
> +				block_unaligned_IO(inode, iocb, iter))
> +		return true;
> +	return false;
>  }
>  
>  #ifdef CONFIG_F2FS_FAULT_INJECTION
> diff --git a/fs/f2fs/file.c b/fs/f2fs/file.c
> index a75f3e145bf1..a388866e71ee 100644
> --- a/fs/f2fs/file.c
> +++ b/fs/f2fs/file.c
> @@ -3019,7 +3019,8 @@ static ssize_t f2fs_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
>  				if (!f2fs_overwrite_io(inode, iocb->ki_pos,
>  						iov_iter_count(from)) ||
>  					f2fs_has_inline_data(inode) ||
> -					f2fs_force_buffered_io(inode, WRITE)) {
> +					f2fs_force_buffered_io(inode,
> +							iocb, from)) {
>  						clear_inode_flag(inode,
>  								FI_NO_PREALLOC);
>  						inode_unlock(inode);
> -- 
> 2.18.0
> 
> 
> 
> _______________________________________________
> Linux-f2fs-devel mailing list
> Linux-f2fs-devel@...ts.sourceforge.net
> https://lists.sourceforge.net/lists/listinfo/linux-f2fs-devel

-- 
--
Sent by a consultant of the Qualcomm Innovation Center, Inc.
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum.

Powered by blists - more mailing lists