Re: [PATCH v5 1/2] f2fs: introduce proc/fs/f2fs/<dev>/fsck_stack node

[Chao Yu <chao@kernel.org>]

On 2021/8/14 2:33, 李扬韬 wrote:
> From: Chao Yu <chao@kernel.org>
> Date: 2021-08-13 22:44:49
> To:  Yangtao Li <frank.li@vivo.com>,jaegeuk@kernel.org
> Cc:  linux-f2fs-devel@lists.sourceforge.net,linux-kernel@vger.kernel.org
> Subject: Re: [PATCH v5 1/2] f2fs: introduce proc/fs/f2fs/<dev>/fsck_stack node>On 2021/8/13 20:32, Yangtao Li wrote:
>>> SBI_NEED_FSCK is an indicator that fsck.f2fs needs to be triggered,
>>> this flag is set in too many places. For some scenes that are not very
>>> reproducible, adding stack information will help locate the problem.
>>>
>>> Let's expose all fsck stack history in procfs.
>>>
>>> Signed-off-by: Yangtao Li <frank.li@vivo.com>
>>> ---
>>> v5:
>>> -fix implicit declaration of function 'stack_trace_save'
>>>    fs/f2fs/f2fs.h  | 34 +++++++++++++++++++++++++++++++++-
>>>    fs/f2fs/sysfs.c | 26 ++++++++++++++++++++++++++
>>>    2 files changed, 59 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/fs/f2fs/f2fs.h b/fs/f2fs/f2fs.h
>>> index 67faa43cc141..cbd06dea3c6a 100644
>>> --- a/fs/f2fs/f2fs.h
>>> +++ b/fs/f2fs/f2fs.h
>>> @@ -24,6 +24,8 @@
>>>    #include <linux/quotaops.h>
>>>    #include <linux/part_stat.h>
>>>    #include <crypto/hash.h>
>>> +#include <linux/stackdepot.h>
>>> +#include <linux/stacktrace.h>
>>>    
>>>    #include <linux/fscrypt.h>
>>>    #include <linux/fsverity.h>
>>> @@ -119,6 +121,8 @@ typedef u32 nid_t;
>>>    
>>>    #define COMPRESS_EXT_NUM		16
>>>    
>>> +#define FSCK_STACK_DEPTH 64
>>> +
>>>    struct f2fs_mount_info {
>>>    	unsigned int opt;
>>>    	int write_io_size_bits;		/* Write IO size bits */
>>> @@ -1786,6 +1790,8 @@ struct f2fs_sb_info {
>>>    	unsigned int compress_watermark;	/* cache page watermark */
>>>    	atomic_t compress_page_hit;		/* cache hit count */
>>>    #endif
>>> +	depot_stack_handle_t *fsck_stack;
>>> +	unsigned int fsck_count;
>>>    };
>>>    
>>>    struct f2fs_private_dio {
>>> @@ -1997,9 +2003,35 @@ static inline bool is_sbi_flag_set(struct f2fs_sb_info *sbi, unsigned int type)
>>>    	return test_bit(type, &sbi->s_flag);
>>>    }
>>>    
>>> -static inline void set_sbi_flag(struct f2fs_sb_info *sbi, unsigned int type)
>>> +static void set_sbi_flag(struct f2fs_sb_info *sbi, unsigned int type)
>>>    {
>>>    	set_bit(type, &sbi->s_flag);
>>> +
>>> +	if (unlikely(type ==  SBI_NEED_FSCK)) {
>>
>> Hmm... I don't know what to say...
> 
> Sorry, maybe there is a problem with my vim configuration.
> The strange thing is that checkpatch.pl didn't check it out.
> 
>>
>>> +		unsigned long entries[FSCK_STACK_DEPTH];
>>> +		depot_stack_handle_t stack, *new;
>>> +		unsigned int nr_entries;
>>> +		int i;
>>> +
>>> +		nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 0);
>>> +		nr_entries = filter_irq_stacks(entries, nr_entries);
>>> +		stack = stack_depot_save(entries, nr_entries, GFP_KERNEL);
>>> +		if (!stack)
>>> +			return;
>>> +
>>> +		/* Try to find an existing entry for this backtrace */
>>> +		for (i = 0; i < sbi->fsck_count; i++)
>>> +			if (sbi->fsck_stack[i] == stack)
>>
>> stack need to be released here?
> 
> We can't remove stack from depot, as we store them contiguously one after
> another in a contiguous memory allocation.
> 
> Or we can limit the recorded stack number.
> 
> $ grep -nr "SBI_NEED_FSCK" fs/f2fs/ --include=*.c --include=*.h | wc -l
> 53
> $ grep -nr "f2fs_bug_on" fs/f2fs/ --include=*.c --include=*.h | wc -l
> 135

I didn't look into details of stack_depot_save(), two stack handles from below
call paths will be the same?

- move_data_block
  - f2fs_wait_on_page_writeback
   - f2fs_bug_on

- ra_data_block
  - f2fs_wait_on_page_writeback
   - f2fs_bug_on

If they have different stack handles, combination number of
set_sbi_flag(NEED_FSCK)/f2fs_bug_on and their callers will be far more than two
hundred.

Thanks,

> 
> Since we only have two hundred possible settings here, considering that
> the same stack will not be recorded, and the probability of occurrence will
> not be high, so it is acceptable not to release?
> 
> If this is the case, the subsequent allocation does not need to be released.
> 
> Thx,
> Yangtao
>