Re: [PATCH v3] f2fs: fix out-of-free problem caused by atomic write

[Chao Yu]

On 2017/11/7 10:24, Jaegeuk Kim wrote:
> On 11/07, Chao Yu wrote:
>> On 2017/11/7 9:55, Jaegeuk Kim wrote:
>>> On 11/06, Yunlong Song wrote:
>>>> f2fs_balance_fs only actives once in the commit_inmem_pages, but there
>>>> are more than one page to commit, so all the other pages will miss the
>>>> check. This will lead to out-of-free problem when commit a very large
>>>> file. However, we cannot do f2fs_balance_fs for each inmem page, since
>>>> this will break atomicity. As a result, we should collect prefree
>>>> segments if needed.
>>>>
>>>> Signed-off-by: Yunlong Song <yunlong.song@xxxxxxxxxx>
>>>> ---
>>>>  fs/f2fs/f2fs.h    |  1 +
>>>>  fs/f2fs/segment.c | 24 +++++++++++++++++++++++-
>>>>  2 files changed, 24 insertions(+), 1 deletion(-)
>>>>
>>>> diff --git a/fs/f2fs/f2fs.h b/fs/f2fs/f2fs.h
>>>> index 13a96b8..04ce48f 100644
>>>> --- a/fs/f2fs/f2fs.h
>>>> +++ b/fs/f2fs/f2fs.h
>>>> @@ -610,6 +610,7 @@ struct f2fs_inode_info {
>>>>  	struct list_head inmem_pages;	/* inmemory pages managed by f2fs */
>>>>  	struct task_struct *inmem_task;	/* store inmemory task */
>>>>  	struct mutex inmem_lock;	/* lock for inmemory pages */
>>>> +	unsigned long inmem_blocks;	/* inmemory blocks */
>>>>  	struct extent_tree *extent_tree;	/* cached extent_tree entry */
>>>>  	struct rw_semaphore dio_rwsem[2];/* avoid racing between dio and gc */
>>>>  	struct rw_semaphore i_mmap_sem;
>>>> diff --git a/fs/f2fs/segment.c b/fs/f2fs/segment.c
>>>> index 46dfbca..b87ede4 100644
>>>> --- a/fs/f2fs/segment.c
>>>> +++ b/fs/f2fs/segment.c
>>>> @@ -210,6 +210,7 @@ void register_inmem_page(struct inode *inode, struct page *page)
>>>>  		list_add_tail(&fi->inmem_ilist, &sbi->inode_list[ATOMIC_FILE]);
>>>>  	spin_unlock(&sbi->inode_lock[ATOMIC_FILE]);
>>>>  	inc_page_count(F2FS_I_SB(inode), F2FS_INMEM_PAGES);
>>>> +	fi->inmem_blocks++;
>>>>  	mutex_unlock(&fi->inmem_lock);
>>>>  
>>>>  	trace_f2fs_register_inmem_page(page, INMEM);
>>>> @@ -221,6 +222,7 @@ static int __revoke_inmem_pages(struct inode *inode,
>>>>  	struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
>>>>  	struct inmem_pages *cur, *tmp;
>>>>  	int err = 0;
>>>> +	struct f2fs_inode_info *fi = F2FS_I(inode);
>>>>  
>>>>  	list_for_each_entry_safe(cur, tmp, head, list) {
>>>>  		struct page *page = cur->page;
>>>> @@ -263,6 +265,7 @@ static int __revoke_inmem_pages(struct inode *inode,
>>>>  		list_del(&cur->list);
>>>>  		kmem_cache_free(inmem_entry_slab, cur);
>>>>  		dec_page_count(F2FS_I_SB(inode), F2FS_INMEM_PAGES);
>>>> +		fi->inmem_blocks--;
>>>>  	}
>>>>  	return err;
>>>>  }
>>>> @@ -302,6 +305,10 @@ void drop_inmem_pages(struct inode *inode)
>>>>  	if (!list_empty(&fi->inmem_ilist))
>>>>  		list_del_init(&fi->inmem_ilist);
>>>>  	spin_unlock(&sbi->inode_lock[ATOMIC_FILE]);
>>>> +	if (fi->inmem_blocks) {
>>>> +		f2fs_bug_on(sbi, 1);
>>>> +		fi->inmem_blocks = 0;
>>>> +	}
>>>>  	mutex_unlock(&fi->inmem_lock);
>>>>  
>>>>  	clear_inode_flag(inode, FI_ATOMIC_FILE);
>>>> @@ -326,6 +333,7 @@ void drop_inmem_page(struct inode *inode, struct page *page)
>>>>  
>>>>  	f2fs_bug_on(sbi, !cur || cur->page != page);
>>>>  	list_del(&cur->list);
>>>> +	fi->inmem_blocks--;
>>>>  	mutex_unlock(&fi->inmem_lock);
>>>>  
>>>>  	dec_page_count(sbi, F2FS_INMEM_PAGES);
>>>> @@ -410,6 +418,16 @@ int commit_inmem_pages(struct inode *inode)
>>>>  
>>>>  	INIT_LIST_HEAD(&revoke_list);
>>>>  	f2fs_balance_fs(sbi, true);
>>>> +	if (prefree_segments(sbi)
>>>> +		&& has_not_enough_free_secs(sbi, 0,
>>>> +		fi->inmem_blocks / BLKS_PER_SEC(sbi))) {
>>>> +		struct cp_control cpc;
>>>> +
>>>> +		cpc.reason = __get_cp_reason(sbi);
>>>> +		err = write_checkpoint(sbi, &cpc);
>>>> +		if (err)
>>>> +			goto drop;
>>>
>>> What do you want to guarantee with this? How about passing target # of segments
>>> into f2fs_balance_fs() so that f2fs_gc() could secure wanted free space in a
>>> loop?
>>
>> Agreed, Jaegeuk, IMO, later we can add one more dirty type F2FS_DIRTY_BUDGET in
>> enum count_type, and introduce below function, add them around dirtying
>> node/dent/imeta datas.
>>
>> void f2fs_budget_space(struct f2fs_sb_info *sbi, unsigned int dirty_budget)
>> {
>> 	if (dirty_budget)
>> 		atomic_add(&sbi->nr_pages[F2FS_DIRTY_BUDGET], dirty_budget);
>>
>> 	f2fs_balance_fs(sbi, dirty_budget);
>> }
>>
>> void f2fs_release_budget(struct f2fs_sb_info *sbi, unsigned int dirty_budget)
>> {
>> 	atomic_dec(&sbi->nr_pages[F2FS_DIRTY_BUDGET], dirty_budget);
>> }
>>
>> So that in has_not_enough_free_secs we can calculate all dirty datas include
>> F2FS_DIRTY_BUDGET type datas more precisely.
>>
>> How about that?
> 
> That'd be actually same as what has_not_enough_free_space() does? Missing part
> would be inmem_pages case which needs quite larger space.

Yup, but I'm not sure all out-of-free segment problems are caused by this. e.g.
concurrent creating may generate lots of dirty nodes which could exceed
f2fs_balance_fs can handle in the end of .create.

Thanks,

> 
>>
>> Thanks,
>>
>>>
>>> Thanks,
>>>
>>>> +	}
>>>>  	f2fs_lock_op(sbi);
>>>>  
>>>>  	set_inode_flag(inode, FI_ATOMIC_COMMIT);
>>>> @@ -429,7 +447,7 @@ int commit_inmem_pages(struct inode *inode)
>>>>  		ret = __revoke_inmem_pages(inode, &revoke_list, false, true);
>>>>  		if (ret)
>>>>  			err = ret;
>>>> -
>>>> +drop:
>>>>  		/* drop all uncommitted pages */
>>>>  		__revoke_inmem_pages(inode, &fi->inmem_pages, true, false);
>>>>  	}
>>>> @@ -437,6 +455,10 @@ int commit_inmem_pages(struct inode *inode)
>>>>  	if (!list_empty(&fi->inmem_ilist))
>>>>  		list_del_init(&fi->inmem_ilist);
>>>>  	spin_unlock(&sbi->inode_lock[ATOMIC_FILE]);
>>>> +	if (fi->inmem_blocks) {
>>>> +		f2fs_bug_on(sbi, 1);
>>>> +		fi->inmem_blocks = 0;
>>>> +	}
>>>>  	mutex_unlock(&fi->inmem_lock);
>>>>  
>>>>  	clear_inode_flag(inode, FI_ATOMIC_COMMIT);
>>>> -- 
>>>> 1.8.5.2
>>>
>>> .
>>>
> 
> .
>