f2fs: fix to map blocks correctly for direct read

[PATCH] f2fs: fix to map blocks correctly for direct read

Posted by Chao Yu 3 weeks, 3 days ago

f2fs_map_blocks() supports to map continuous holes or preallocated
address, we should avoid setting F2FS_MAP_MAPPED for these cases
only, otherwise, it may fail f2fs_iomap_begin(), and make direct
write fallbacking to use buffered IO and flush, result in performance
regression.

Fixes: 9f0f6bf42714 ("f2fs: support to map continuous holes or preallocated address")
Reported-by: kernel test robot <oliver.sang@intel.com>
Closes: https://lore.kernel.org/oe-lkp/202409122103.e45aa13b-oliver.sang@intel.com
Cc: Cyril Hrubis <chrubis@suse.cz>
Signed-off-by: Chao Yu <chao@kernel.org>
---
 fs/f2fs/data.c | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)

diff --git a/fs/f2fs/data.c b/fs/f2fs/data.c
index 306b86b0595d..38b85160c6bd 100644
--- a/fs/f2fs/data.c
+++ b/fs/f2fs/data.c
@@ -1676,7 +1676,12 @@ int f2fs_map_blocks(struct inode *inode, struct f2fs_map_blocks *map, int flag)
 		/* reserved delalloc block should be mapped for fiemap. */
 		if (blkaddr == NEW_ADDR)
 			map->m_flags |= F2FS_MAP_DELALLOC;
-		if (flag != F2FS_GET_BLOCK_DIO || !is_hole)
+		/*
+		 * f2fs_map_blocks() supports to map continuous holes or
+		 * preallocated address, don't set F2FS_MAP_MAPPED for these
+		 * cases only.
+		 */
+		if (flag != F2FS_GET_BLOCK_DIO || map->m_may_create || !is_hole)
 			map->m_flags |= F2FS_MAP_MAPPED;
 
 		map->m_pblk = blkaddr;
-- 
2.40.1

Re: [PATCH] f2fs: fix to map blocks correctly for direct read

Posted by Jaegeuk Kim 3 weeks, 2 days ago

On 10/31, Chao Yu wrote:
> f2fs_map_blocks() supports to map continuous holes or preallocated
> address, we should avoid setting F2FS_MAP_MAPPED for these cases
> only, otherwise, it may fail f2fs_iomap_begin(), and make direct
> write fallbacking to use buffered IO and flush, result in performance
> regression.

Is this fixing direct write or read?

> 
> Fixes: 9f0f6bf42714 ("f2fs: support to map continuous holes or preallocated address")
> Reported-by: kernel test robot <oliver.sang@intel.com>
> Closes: https://lore.kernel.org/oe-lkp/202409122103.e45aa13b-oliver.sang@intel.com
> Cc: Cyril Hrubis <chrubis@suse.cz>
> Signed-off-by: Chao Yu <chao@kernel.org>
> ---
>  fs/f2fs/data.c | 7 ++++++-
>  1 file changed, 6 insertions(+), 1 deletion(-)
> 
> diff --git a/fs/f2fs/data.c b/fs/f2fs/data.c
> index 306b86b0595d..38b85160c6bd 100644
> --- a/fs/f2fs/data.c
> +++ b/fs/f2fs/data.c
> @@ -1676,7 +1676,12 @@ int f2fs_map_blocks(struct inode *inode, struct f2fs_map_blocks *map, int flag)
>  		/* reserved delalloc block should be mapped for fiemap. */
>  		if (blkaddr == NEW_ADDR)
>  			map->m_flags |= F2FS_MAP_DELALLOC;
> -		if (flag != F2FS_GET_BLOCK_DIO || !is_hole)
> +		/*
> +		 * f2fs_map_blocks() supports to map continuous holes or
> +		 * preallocated address, don't set F2FS_MAP_MAPPED for these
> +		 * cases only.
> +		 */
> +		if (flag != F2FS_GET_BLOCK_DIO || map->m_may_create || !is_hole)

So, this is adding map->m_may_create, which is for writes?

>  			map->m_flags |= F2FS_MAP_MAPPED;
>  
>  		map->m_pblk = blkaddr;
> -- 
> 2.40.1

Re: [PATCH] f2fs: fix to map blocks correctly for direct read

Posted by Chao Yu 3 weeks, 2 days ago

On 2024/11/1 9:22, Jaegeuk Kim wrote:
> On 10/31, Chao Yu wrote:
>> f2fs_map_blocks() supports to map continuous holes or preallocated
>> address, we should avoid setting F2FS_MAP_MAPPED for these cases
>> only, otherwise, it may fail f2fs_iomap_begin(), and make direct
>> write fallbacking to use buffered IO and flush, result in performance
>> regression.
> 
> Is this fixing direct write or read?

Direct write.

> 
>>
>> Fixes: 9f0f6bf42714 ("f2fs: support to map continuous holes or preallocated address")
>> Reported-by: kernel test robot <oliver.sang@intel.com>
>> Closes: https://lore.kernel.org/oe-lkp/202409122103.e45aa13b-oliver.sang@intel.com
>> Cc: Cyril Hrubis <chrubis@suse.cz>
>> Signed-off-by: Chao Yu <chao@kernel.org>
>> ---
>>   fs/f2fs/data.c | 7 ++++++-
>>   1 file changed, 6 insertions(+), 1 deletion(-)
>>
>> diff --git a/fs/f2fs/data.c b/fs/f2fs/data.c
>> index 306b86b0595d..38b85160c6bd 100644
>> --- a/fs/f2fs/data.c
>> +++ b/fs/f2fs/data.c
>> @@ -1676,7 +1676,12 @@ int f2fs_map_blocks(struct inode *inode, struct f2fs_map_blocks *map, int flag)
>>   		/* reserved delalloc block should be mapped for fiemap. */
>>   		if (blkaddr == NEW_ADDR)
>>   			map->m_flags |= F2FS_MAP_DELALLOC;
>> -		if (flag != F2FS_GET_BLOCK_DIO || !is_hole)
>> +		/*
>> +		 * f2fs_map_blocks() supports to map continuous holes or
>> +		 * preallocated address, don't set F2FS_MAP_MAPPED for these
>> +		 * cases only.
>> +		 */
>> +		if (flag != F2FS_GET_BLOCK_DIO || map->m_may_create || !is_hole)
> 
> So, this is adding map->m_may_create, which is for writes?

map->m_may_create is added for write path, w/o this condition, f2fs_map_blocks()
may missed to tag F2FS_MAP_MAPPED, result in that f2fs_iomap_begin() will return
-ENOTBLK.

Am I missing someting?

Thanks,

> 
>>   			map->m_flags |= F2FS_MAP_MAPPED;
>>   
>>   		map->m_pblk = blkaddr;
>> -- 
>> 2.40.1

Re: [PATCH] f2fs: fix to map blocks correctly for direct read

Posted by Jaegeuk Kim 3 weeks, 2 days ago

On 11/01, Chao Yu wrote:
> On 2024/11/1 9:22, Jaegeuk Kim wrote:
> > On 10/31, Chao Yu wrote:
> > > f2fs_map_blocks() supports to map continuous holes or preallocated
> > > address, we should avoid setting F2FS_MAP_MAPPED for these cases
> > > only, otherwise, it may fail f2fs_iomap_begin(), and make direct
> > > write fallbacking to use buffered IO and flush, result in performance
> > > regression.
> > 
> > Is this fixing direct write or read?
> 
> Direct write.

I was confising the patch subject is saying direct read.

> 
> > 
> > > 
> > > Fixes: 9f0f6bf42714 ("f2fs: support to map continuous holes or preallocated address")
> > > Reported-by: kernel test robot <oliver.sang@intel.com>
> > > Closes: https://lore.kernel.org/oe-lkp/202409122103.e45aa13b-oliver.sang@intel.com
> > > Cc: Cyril Hrubis <chrubis@suse.cz>
> > > Signed-off-by: Chao Yu <chao@kernel.org>
> > > ---
> > >   fs/f2fs/data.c | 7 ++++++-
> > >   1 file changed, 6 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/fs/f2fs/data.c b/fs/f2fs/data.c
> > > index 306b86b0595d..38b85160c6bd 100644
> > > --- a/fs/f2fs/data.c
> > > +++ b/fs/f2fs/data.c
> > > @@ -1676,7 +1676,12 @@ int f2fs_map_blocks(struct inode *inode, struct f2fs_map_blocks *map, int flag)
> > >   		/* reserved delalloc block should be mapped for fiemap. */
> > >   		if (blkaddr == NEW_ADDR)
> > >   			map->m_flags |= F2FS_MAP_DELALLOC;
> > > -		if (flag != F2FS_GET_BLOCK_DIO || !is_hole)
> > > +		/*
> > > +		 * f2fs_map_blocks() supports to map continuous holes or
> > > +		 * preallocated address, don't set F2FS_MAP_MAPPED for these
> > > +		 * cases only.
> > > +		 */
> > > +		if (flag != F2FS_GET_BLOCK_DIO || map->m_may_create || !is_hole)
> > 
> > So, this is adding map->m_may_create, which is for writes?
> 
> map->m_may_create is added for write path, w/o this condition, f2fs_map_blocks()
> may missed to tag F2FS_MAP_MAPPED, result in that f2fs_iomap_begin() will return
> -ENOTBLK.

Is that something like this?

/* DIO READ and hole case, should not map the blocks. */
if (!(flag == F2FS_GET_BLOCK_DIO && is_hole && !map->m_may_create))

> 
> Am I missing someting?
> 
> Thanks,
> 
> > 
> > >   			map->m_flags |= F2FS_MAP_MAPPED;
> > >   		map->m_pblk = blkaddr;
> > > -- 
> > > 2.40.1

Re: [PATCH] f2fs: fix to map blocks correctly for direct read

Posted by Chao Yu 3 weeks, 1 day ago

On 2024/11/2 5:06, Jaegeuk Kim wrote:
> On 11/01, Chao Yu wrote:
>> On 2024/11/1 9:22, Jaegeuk Kim wrote:
>>> On 10/31, Chao Yu wrote:
>>>> f2fs_map_blocks() supports to map continuous holes or preallocated
>>>> address, we should avoid setting F2FS_MAP_MAPPED for these cases
>>>> only, otherwise, it may fail f2fs_iomap_begin(), and make direct
>>>> write fallbacking to use buffered IO and flush, result in performance
>>>> regression.
>>>
>>> Is this fixing direct write or read?
>>
>> Direct write.
> 
> I was confising the patch subject is saying direct read.

Oops, sorry.

> 
>>
>>>
>>>>
>>>> Fixes: 9f0f6bf42714 ("f2fs: support to map continuous holes or preallocated address")
>>>> Reported-by: kernel test robot <oliver.sang@intel.com>
>>>> Closes: https://lore.kernel.org/oe-lkp/202409122103.e45aa13b-oliver.sang@intel.com
>>>> Cc: Cyril Hrubis <chrubis@suse.cz>
>>>> Signed-off-by: Chao Yu <chao@kernel.org>
>>>> ---
>>>>    fs/f2fs/data.c | 7 ++++++-
>>>>    1 file changed, 6 insertions(+), 1 deletion(-)
>>>>
>>>> diff --git a/fs/f2fs/data.c b/fs/f2fs/data.c
>>>> index 306b86b0595d..38b85160c6bd 100644
>>>> --- a/fs/f2fs/data.c
>>>> +++ b/fs/f2fs/data.c
>>>> @@ -1676,7 +1676,12 @@ int f2fs_map_blocks(struct inode *inode, struct f2fs_map_blocks *map, int flag)
>>>>    		/* reserved delalloc block should be mapped for fiemap. */
>>>>    		if (blkaddr == NEW_ADDR)
>>>>    			map->m_flags |= F2FS_MAP_DELALLOC;
>>>> -		if (flag != F2FS_GET_BLOCK_DIO || !is_hole)
>>>> +		/*
>>>> +		 * f2fs_map_blocks() supports to map continuous holes or
>>>> +		 * preallocated address, don't set F2FS_MAP_MAPPED for these
>>>> +		 * cases only.
>>>> +		 */
>>>> +		if (flag != F2FS_GET_BLOCK_DIO || map->m_may_create || !is_hole)
>>>
>>> So, this is adding map->m_may_create, which is for writes?
>>
>> map->m_may_create is added for write path, w/o this condition, f2fs_map_blocks()
>> may missed to tag F2FS_MAP_MAPPED, result in that f2fs_iomap_begin() will return
>> -ENOTBLK.
> 
> Is that something like this?
> 
> /* DIO READ and hole case, should not map the blocks. */
> if (!(flag == F2FS_GET_BLOCK_DIO && is_hole && !map->m_may_create))

More clear, will update w/ this.

Thanks,

> 
>>
>> Am I missing someting?
>>
>> Thanks,
>>
>>>
>>>>    			map->m_flags |= F2FS_MAP_MAPPED;
>>>>    		map->m_pblk = blkaddr;
>>>> -- 
>>>> 2.40.1