2021-04-14 11:24:28

by Chao Yu

[permalink] [raw]
Subject: [PATCH] f2fs: fix to cover allocate_segment() with lock

As we did for other cases, in fix_curseg_write_pointer(), let's
change as below:
- use callback function s_ops->allocate_segment() instead of
raw function allocate_segment_by_default();
- cover allocate_segment() with curseg_lock and sentry_lock.

Signed-off-by: Chao Yu <[email protected]>
---
fs/f2fs/segment.c | 7 ++++++-
1 file changed, 6 insertions(+), 1 deletion(-)

diff --git a/fs/f2fs/segment.c b/fs/f2fs/segment.c
index b2ee6b7791b0..daf9531ec58f 100644
--- a/fs/f2fs/segment.c
+++ b/fs/f2fs/segment.c
@@ -4848,7 +4848,12 @@ static int fix_curseg_write_pointer(struct f2fs_sb_info *sbi, int type)

f2fs_notice(sbi, "Assign new section to curseg[%d]: "
"curseg[0x%x,0x%x]", type, cs->segno, cs->next_blkoff);
- allocate_segment_by_default(sbi, type, true);
+
+ down_read(&SM_I(sbi)->curseg_lock);
+ down_write(&SIT_I(sbi)->sentry_lock);
+ SIT_I(sbi)->s_ops->allocate_segment(sbi, type, true);
+ up_write(&SIT_I(sbi)->sentry_lock);
+ up_read(&SM_I(sbi)->curseg_lock);

/* check consistency of the zone curseg pointed to */
if (check_zone_write_pointer(sbi, zbd, &zone))
--
2.29.2


2021-04-19 16:59:00

by Jaegeuk Kim

[permalink] [raw]
Subject: Re: [PATCH] f2fs: fix to cover allocate_segment() with lock

On 04/14, Chao Yu wrote:
> As we did for other cases, in fix_curseg_write_pointer(), let's
> change as below:
> - use callback function s_ops->allocate_segment() instead of
> raw function allocate_segment_by_default();
> - cover allocate_segment() with curseg_lock and sentry_lock.
>
> Signed-off-by: Chao Yu <[email protected]>
> ---
> fs/f2fs/segment.c | 7 ++++++-
> 1 file changed, 6 insertions(+), 1 deletion(-)
>
> diff --git a/fs/f2fs/segment.c b/fs/f2fs/segment.c
> index b2ee6b7791b0..daf9531ec58f 100644
> --- a/fs/f2fs/segment.c
> +++ b/fs/f2fs/segment.c
> @@ -4848,7 +4848,12 @@ static int fix_curseg_write_pointer(struct f2fs_sb_info *sbi, int type)
>
> f2fs_notice(sbi, "Assign new section to curseg[%d]: "
> "curseg[0x%x,0x%x]", type, cs->segno, cs->next_blkoff);
> - allocate_segment_by_default(sbi, type, true);
> +
> + down_read(&SM_I(sbi)->curseg_lock);
> + down_write(&SIT_I(sbi)->sentry_lock);
> + SIT_I(sbi)->s_ops->allocate_segment(sbi, type, true);
> + up_write(&SIT_I(sbi)->sentry_lock);
> + up_read(&SM_I(sbi)->curseg_lock);

Seems f2fs_allocate_new_section()?

>
> /* check consistency of the zone curseg pointed to */
> if (check_zone_write_pointer(sbi, zbd, &zone))
> --
> 2.29.2

2021-04-20 01:54:06

by Chao Yu

[permalink] [raw]
Subject: Re: [PATCH] f2fs: fix to cover allocate_segment() with lock

On 2021/4/20 0:57, Jaegeuk Kim wrote:
> On 04/14, Chao Yu wrote:
>> As we did for other cases, in fix_curseg_write_pointer(), let's
>> change as below:
>> - use callback function s_ops->allocate_segment() instead of
>> raw function allocate_segment_by_default();
>> - cover allocate_segment() with curseg_lock and sentry_lock.
>>
>> Signed-off-by: Chao Yu <[email protected]>
>> ---
>> fs/f2fs/segment.c | 7 ++++++-
>> 1 file changed, 6 insertions(+), 1 deletion(-)
>>
>> diff --git a/fs/f2fs/segment.c b/fs/f2fs/segment.c
>> index b2ee6b7791b0..daf9531ec58f 100644
>> --- a/fs/f2fs/segment.c
>> +++ b/fs/f2fs/segment.c
>> @@ -4848,7 +4848,12 @@ static int fix_curseg_write_pointer(struct f2fs_sb_info *sbi, int type)
>>
>> f2fs_notice(sbi, "Assign new section to curseg[%d]: "
>> "curseg[0x%x,0x%x]", type, cs->segno, cs->next_blkoff);
>> - allocate_segment_by_default(sbi, type, true);
>> +
>> + down_read(&SM_I(sbi)->curseg_lock);
>> + down_write(&SIT_I(sbi)->sentry_lock);
>> + SIT_I(sbi)->s_ops->allocate_segment(sbi, type, true);
>> + up_write(&SIT_I(sbi)->sentry_lock);
>> + up_read(&SM_I(sbi)->curseg_lock);
>
> Seems f2fs_allocate_new_section()?

f2fs_allocate_new_section() will allocate new section only when current
section has been initialized and has valid block/ckpt_block.

It looks fix_curseg_write_pointer() wants to force migrating current segment
to new section whenever write pointer and curseg->next_blkoff is inconsistent.

So how about adding a parameter to force f2fs_allocate_new_section() to
allocate new section?

Thanks,

>
>>
>> /* check consistency of the zone curseg pointed to */
>> if (check_zone_write_pointer(sbi, zbd, &zone))
>> --
>> 2.29.2
> .
>

2021-04-20 18:26:58

by Jaegeuk Kim

[permalink] [raw]
Subject: Re: [PATCH] f2fs: fix to cover allocate_segment() with lock

On 04/20, Chao Yu wrote:
> On 2021/4/20 0:57, Jaegeuk Kim wrote:
> > On 04/14, Chao Yu wrote:
> > > As we did for other cases, in fix_curseg_write_pointer(), let's
> > > change as below:
> > > - use callback function s_ops->allocate_segment() instead of
> > > raw function allocate_segment_by_default();
> > > - cover allocate_segment() with curseg_lock and sentry_lock.
> > >
> > > Signed-off-by: Chao Yu <[email protected]>
> > > ---
> > > fs/f2fs/segment.c | 7 ++++++-
> > > 1 file changed, 6 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/fs/f2fs/segment.c b/fs/f2fs/segment.c
> > > index b2ee6b7791b0..daf9531ec58f 100644
> > > --- a/fs/f2fs/segment.c
> > > +++ b/fs/f2fs/segment.c
> > > @@ -4848,7 +4848,12 @@ static int fix_curseg_write_pointer(struct f2fs_sb_info *sbi, int type)
> > > f2fs_notice(sbi, "Assign new section to curseg[%d]: "
> > > "curseg[0x%x,0x%x]", type, cs->segno, cs->next_blkoff);
> > > - allocate_segment_by_default(sbi, type, true);
> > > +
> > > + down_read(&SM_I(sbi)->curseg_lock);
> > > + down_write(&SIT_I(sbi)->sentry_lock);
> > > + SIT_I(sbi)->s_ops->allocate_segment(sbi, type, true);
> > > + up_write(&SIT_I(sbi)->sentry_lock);
> > > + up_read(&SM_I(sbi)->curseg_lock);
> >
> > Seems f2fs_allocate_new_section()?
>
> f2fs_allocate_new_section() will allocate new section only when current
> section has been initialized and has valid block/ckpt_block.
>
> It looks fix_curseg_write_pointer() wants to force migrating current segment
> to new section whenever write pointer and curseg->next_blkoff is inconsistent.
>
> So how about adding a parameter to force f2fs_allocate_new_section() to
> allocate new section?

I think that can be doable. Hope to avoid native calls as much as possible.

>
> Thanks,
>
> >
> > > /* check consistency of the zone curseg pointed to */
> > > if (check_zone_write_pointer(sbi, zbd, &zone))
> > > --
> > > 2.29.2
> > .
> >