Re: [PATCH] fat: Batched discard support for fat

[Lukas Czerner]

On Thu, 17 Feb 2011, Kyungmin Park wrote:

> From: Kyungmin Park <kyungmin.park@xxxxxxxxxxx>
> 
> FAT supports batched discard as ext4.
> 
> Cited from Lukas words.
> "The current solution is not ideal because of its bad performance impact.
> So basic idea to improve things is to avoid discarding every time some
> blocks are freed. and instead batching is together into bigger trims,
> which tends to be more effective."
> 
> You can find an information in detail at following URLs.
> http://lwn.net/Articles/397538/
> http://lwn.net/Articles/383933/

Hi Kyungmin,

this is really great to see more and more filesystemS ADDING this. I can
not really comment on fat specific code, but anyway I have couple of
comments bellow.

Thanks!
-Lukas

> 
> Signed-off-by: Kyungmin Park <kyungmin.park@xxxxxxxxxxx>
> ---
> diff --git a/fs/fat/fat.h b/fs/fat/fat.h
> index f504089..08b53e1 100644
> --- a/fs/fat/fat.h
> +++ b/fs/fat/fat.h
> @@ -299,6 +299,7 @@ extern int fat_alloc_clusters(struct inode *inode, int *cluster,
>  			      int nr_cluster);
>  extern int fat_free_clusters(struct inode *inode, int cluster);
>  extern int fat_count_free_clusters(struct super_block *sb);
> +extern int fat_trim_fs(struct super_block *sb, struct fstrim_range *range);
>  
>  /* fat/file.c */
>  extern long fat_generic_ioctl(struct file *filp, unsigned int cmd,
> diff --git a/fs/fat/fatent.c b/fs/fat/fatent.c
> index b47d2c9..777094b 100644
> --- a/fs/fat/fatent.c
> +++ b/fs/fat/fatent.c
> @@ -1,6 +1,8 @@
>  /*
>   * Copyright (C) 2004, OGAWA Hirofumi
>   * Released under GPL v2.
> + *
> + * Batched discard support by Kyungmin Park <kyungmin.park@xxxxxxxxxxx>
>   */
>  
>  #include <linux/module.h>
> @@ -541,6 +543,16 @@ out:
>  	return err;
>  }
>  
> +static int fat_issue_discard(struct super_block *sb, int cluster, int nr_clus)
> +{
> +	struct msdos_sb_info *sbi = MSDOS_SB(sb);
> +	sector_t block, nr_blocks;
> +
> +        block = fat_clus_to_blknr(sbi, cluster);
> +        nr_blocks = nr_clus * sbi->sec_per_clus;
> +        return sb_issue_discard(sb, block, nr_blocks, GFP_NOFS, 0);

Use tabs for code indent.

> +}
> +
>  int fat_free_clusters(struct inode *inode, int cluster)
>  {
>  	struct super_block *sb = inode->i_sb;
> @@ -575,11 +587,7 @@ int fat_free_clusters(struct inode *inode, int cluster)
>  			if (cluster != fatent.entry + 1) {
>  				int nr_clus = fatent.entry - first_cl + 1;
>  
> -				sb_issue_discard(sb,
> -					fat_clus_to_blknr(sbi, first_cl),
> -					nr_clus * sbi->sec_per_clus,
> -					GFP_NOFS, 0);
> -
> +				fat_issue_discard(sb, first_cl, nr_clus);
>  				first_cl = cluster;
>  			}
>  		}
> @@ -683,3 +691,73 @@ out:
>  	unlock_fat(sbi);
>  	return err;
>  }
> +
> +int fat_trim_fs(struct super_block *sb, struct fstrim_range *range)
> +{
> +	struct msdos_sb_info *sbi = MSDOS_SB(sb);
> +	struct fatent_operations *ops = sbi->fatent_ops;
> +	struct fat_entry fatent;
> +	unsigned long reada_blocks, reada_mask, cur_block;
> +	int err = 0, free;
> +	unsigned int start, len, minlen, trimmed;
> +	int entry = 0;
> +
> +	start = range->start >> sb->s_blocksize_bits;

Start is not used anywhere either. Let me just explain the reason for
having start, and len.

start - address the first Byte of the filesystem (or better, the device,
  or part of the device underneath the filesystem itself). You are
  supposed to use this as a starting point for doing discard.

len - defines the length in Bytes of the filesystem (or better, the
  device, or part of the device underneath the filesystem itself), user
  want to discard from the "start".

This two values gives us more flexibility in the way we are discarding
the filesystem's free blocks. For example, if you have huge filesystem,
or your device has bad discard performance, you probably do not want to
do the FITRIM all-at-once, but rather per-partes, to not disturb other
ongoing IO too much. Basically it allows you to spread the load through
longer period of time.

Also, for bigger filesystems we might want to inform user that something
is really happening in the form of progress bar (or whatever), which you
can not do otherwise.

> +	len = range->len >> sb->s_blocksize_bits;

"len" does not seem to be used anywhere, does it mean that you are
discarding free extents on the whole filesystem with one run ? This
operation can take pretty long time on devices with slow discard
capability. Can you consider doing it per-partes as we are doing in
ext4 for example ? I really can't say how much work does it mean in
fat, but it might be worth it. Especially because of ... ->

> +	minlen = range->minlen >> sb->s_blocksize_bits;
> +	trimmed = 0;
> +
> +	minlen = minlen / sbi->sec_per_clus;
> +
> +	lock_fat(sbi);
	^^^^^^^^^^^^^^
	-> this
You are holding this mutex the whole time you are discarding free
extents and I think it would be devastating, because the users will see
long stalls. The bigger the filesystem is and the worse discard
performance of the device is, the longer stall will be. Please correct
me I am mistaken. Have you done any testing ?

> +	if (sbi->free_clusters != -1 && sbi->free_clus_valid)
> +		goto out;
> +
> +	reada_blocks = FAT_READA_SIZE >> sb->s_blocksize_bits;
> +	reada_mask = reada_blocks - 1;
> +	cur_block = 0;
> +
> +	free = 0;
> +	fatent_init(&fatent);
> +
> +	/*
> +	 * REVISIT: scan from the last free block.
> +	 */
> +	fatent_set_entry(&fatent, FAT_START_ENT);
> +	while (fatent.entry < sbi->max_cluster) {
> +		/* readahead of fat blocks */
> +		if ((cur_block & reada_mask) == 0) {
> +			unsigned long rest = sbi->fat_length - cur_block;
> +			fat_ent_reada(sb, &fatent, min(reada_blocks, rest));
> +		}
> +		cur_block++;
> +
> +		err = fat_ent_read_block(sb, &fatent);
> +		if (err)
> +			goto out;
> +
> +		do {
> +			if (ops->ent_get(&fatent) == FAT_ENT_FREE) {
> +				free++;
> +				if (!entry)
> +					entry = fatent.entry;
> +			} else if (entry) {
> +				if (free >= minlen) {
> +					fat_issue_discard(sb, entry, free);
> +					trimmed += free;
> +				}
> +				free = 0;
> +				entry = 0;
> +			}
> +		} while (fat_ent_next(sbi, &fatent));
> +	}
> +	if (free >= minlen) {
> +		fat_issue_discard(sb, entry, free);
> +		trimmed += free;
> +	}
> +	range->len = (trimmed * sbi->sec_per_clus) * sb->s_blocksize;
> +	fatent_brelse(&fatent);
> +out:
> +	unlock_fat(sbi);
> +	return err;
> +}
> diff --git a/fs/fat/file.c b/fs/fat/file.c
> index 7257752..bfdd558 100644
> --- a/fs/fat/file.c
> +++ b/fs/fat/file.c
> @@ -125,6 +125,30 @@ long fat_generic_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
>  		return fat_ioctl_get_attributes(inode, user_attr);
>  	case FAT_IOCTL_SET_ATTRIBUTES:
>  		return fat_ioctl_set_attributes(filp, user_attr);
> +	case FITRIM:
> +	{
> +		struct super_block *sb = inode->i_sb;
> +		struct fstrim_range range;
> +		int ret = 0;
> +
> +		if (!capable(CAP_SYS_ADMIN))
> +			return -EPERM;

You might want to add check whether the device actually support discard.
See: http://www.spinics.net/lists/linux-ext4/msg23144.html

> +
> +		if (copy_from_user(&range, (struct fstrim_range *)arg,
> +					sizeof(range)))
> +			return -EFAULT;
> +
> +		ret = fat_trim_fs(sb, &range);
> +		if (ret < 0)
> +			return ret;
> +
> +		if (copy_to_user((struct fstrim_range *)arg, &range,
> +					sizeof(range)))
> +			return -EFAULT;
> +
> +		return 0;
> +	}
> +
>  	default:
>  		return -ENOTTY;	/* Inappropriate ioctl for device */
>  	}
> 

-- 
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/