Re: [Ocfs2-devel] [PATCH 3/3] ocfs2: nowait aio support

[Gang He]

Hello Alex,


>>> 
> Hi Gang,
> 
> On 2017/11/27 17:46, Gang He wrote:
>> Return EAGAIN if any of the following checks fail for direct I/O:
>> Can not get the related locks immediately,
>> Blocks are not allocated at the write location, it will trigger
>> block allocation and block IO operations.
>> 
>> Signed-off-by: Gang He <ghe@xxxxxxxx>
>> ---
>>  fs/ocfs2/dir.c         |  2 +-
>>  fs/ocfs2/dlmglue.c     | 20 ++++++++++----
>>  fs/ocfs2/dlmglue.h     |  2 +-
>>  fs/ocfs2/file.c        | 74 +++++++++++++++++++++++++++++++++++++-------------
>>  fs/ocfs2/mmap.c        |  2 +-
>>  fs/ocfs2/ocfs2_trace.h | 10 ++++---
>>  6 files changed, 79 insertions(+), 31 deletions(-)
>> 
>> diff --git a/fs/ocfs2/dir.c b/fs/ocfs2/dir.c
>> index febe631..ea50901 100644
>> --- a/fs/ocfs2/dir.c
>> +++ b/fs/ocfs2/dir.c
>> @@ -1957,7 +1957,7 @@ int ocfs2_readdir(struct file *file, struct dir_context 
> *ctx)
>>  
>>  	trace_ocfs2_readdir((unsigned long long)OCFS2_I(inode)->ip_blkno);
>>  
>> -	error = ocfs2_inode_lock_atime(inode, file->f_path.mnt, &lock_level);
>> +	error = ocfs2_inode_lock_atime(inode, file->f_path.mnt, &lock_level, 1);
>>  	if (lock_level && error >= 0) {
>>  		/* We release EX lock which used to update atime
>>  		 * and get PR lock again to reduce contention
>> diff --git a/fs/ocfs2/dlmglue.c b/fs/ocfs2/dlmglue.c
>> index 5cfbd04..feb8dbe 100644
>> --- a/fs/ocfs2/dlmglue.c
>> +++ b/fs/ocfs2/dlmglue.c
>> @@ -2516,13 +2516,18 @@ int ocfs2_inode_lock_with_page(struct inode *inode,
>>  
>>  int ocfs2_inode_lock_atime(struct inode *inode,
>>  			  struct vfsmount *vfsmnt,
>> -			  int *level)
>> +			  int *level, int wait)
>>  {
>>  	int ret;
>>  
>> -	ret = ocfs2_inode_lock(inode, NULL, 0);
>> +	if (wait)
>> +		ret = ocfs2_inode_lock(inode, NULL, 0);
>> +	else
>> +		ret = ocfs2_try_inode_lock(inode, NULL, 0);
>> +
>>  	if (ret < 0) {
>> -		mlog_errno(ret);
>> +		if (ret != -EAGAIN)
>> +			mlog_errno(ret);
>>  		return ret;
>>  	}
>>  
>> @@ -2534,9 +2539,14 @@ int ocfs2_inode_lock_atime(struct inode *inode,
>>  		struct buffer_head *bh = NULL;
>>  
>>  		ocfs2_inode_unlock(inode, 0);
>> -		ret = ocfs2_inode_lock(inode, &bh, 1);
>> +		if (wait)
>> +			ret = ocfs2_inode_lock(inode, &bh, 1);
>> +		else
>> +			ret = ocfs2_try_inode_lock(inode, &bh, 1);
>> +
>>  		if (ret < 0) {
>> -			mlog_errno(ret);
>> +			if (ret != -EAGAIN)
>> +				mlog_errno(ret);
>>  			return ret;
>>  		}
>>  		*level = 1;
>> diff --git a/fs/ocfs2/dlmglue.h b/fs/ocfs2/dlmglue.h
>> index 05910fc..c83dbb5 100644
>> --- a/fs/ocfs2/dlmglue.h
>> +++ b/fs/ocfs2/dlmglue.h
>> @@ -123,7 +123,7 @@ void ocfs2_refcount_lock_res_init(struct ocfs2_lock_res 
> *lockres,
>>  void ocfs2_open_unlock(struct inode *inode);
>>  int ocfs2_inode_lock_atime(struct inode *inode,
>>  			  struct vfsmount *vfsmnt,
>> -			  int *level);
>> +			  int *level, int wait);
>>  int ocfs2_inode_lock_full_nested(struct inode *inode,
>>  			 struct buffer_head **ret_bh,
>>  			 int ex,
>> diff --git a/fs/ocfs2/file.c b/fs/ocfs2/file.c
>> index dc455d4..900f04e 100644
>> --- a/fs/ocfs2/file.c
>> +++ b/fs/ocfs2/file.c
>> @@ -140,6 +140,8 @@ static int ocfs2_file_open(struct inode *inode, struct 
> file *file)
>>  		spin_unlock(&oi->ip_lock);
>>  	}
>>  
>> +	file->f_mode |= FMODE_NOWAIT;
>> +
>>  leave:
>>  	return status;
>>  }
>> @@ -2132,8 +2134,7 @@ static int ocfs2_prepare_inode_for_refcount(struct 
> inode *inode,
>>  }
>>  
>>  static int ocfs2_prepare_inode_for_write(struct file *file,
>> -					 loff_t pos,
>> -					 size_t count)
>> +					 loff_t pos, size_t count, int wait)
>>  {
>>  	int ret = 0, meta_level = 0;
>>  	struct dentry *dentry = file->f_path.dentry;
>> @@ -2145,10 +2146,14 @@ static int ocfs2_prepare_inode_for_write(struct file 
> *file,
>>  	 * if we need to make modifications here.
>>  	 */
>>  	for(;;) {
>> -		ret = ocfs2_inode_lock(inode, NULL, meta_level);
>> +		if (wait)
>> +			ret = ocfs2_inode_lock(inode, NULL, meta_level);
>> +		else
>> +			ret = ocfs2_try_inode_lock(inode, NULL, meta_level);
>>  		if (ret < 0) {
>>  			meta_level = -1;
>> -			mlog_errno(ret);
>> +			if (ret != -EAGAIN)
>> +				mlog_errno(ret);
>>  			goto out;
>>  		}
>>
> 
> We will lock inode again in 
> ocfs2_prepare_inode_for_write()->ocfs2_prepare_inode_for_refcount().
> Should we add the check of 'nowait' flags?
I think ocfs2_overwrite_io() function can make sure ocfs2_prepare_inode_for_refcount() is passed,
but it looks there is a race condition between ocfs2_overwrite_io() and ocfs2_prepare_inode_for_write() since inode lock is released.
I think I will move ocfs2_overwrite_io() function invoking in ocfs2_prepare_inode_for_write() to avoid this race.

Thanks
Gang


> 
>> @@ -2199,7 +2204,7 @@ static int ocfs2_prepare_inode_for_write(struct file 
> *file,
>>  
>>  out_unlock:
>>  	trace_ocfs2_prepare_inode_for_write(OCFS2_I(inode)->ip_blkno,
>> -					    pos, count);
>> +					    pos, count, wait);
>>  
>>  	if (meta_level >= 0)
>>  		ocfs2_inode_unlock(inode, meta_level);
>> @@ -2211,7 +2216,7 @@ static int ocfs2_prepare_inode_for_write(struct file 
> *file,
>>  static ssize_t ocfs2_file_write_iter(struct kiocb *iocb,
>>  				    struct iov_iter *from)
>>  {
>> -	int direct_io, rw_level;
>> +	int rw_level;
>>  	ssize_t written = 0;
>>  	ssize_t ret;
>>  	size_t count = iov_iter_count(from);
>> @@ -2223,6 +2228,8 @@ static ssize_t ocfs2_file_write_iter(struct kiocb 
> *iocb,
>>  	void *saved_ki_complete = NULL;
>>  	int append_write = ((iocb->ki_pos + count) >=
>>  			i_size_read(inode) ? 1 : 0);
>> +	int direct_io = iocb->ki_flags & IOCB_DIRECT ? 1 : 0;
>> +	int nowait = iocb->ki_flags & IOCB_NOWAIT ? 1 : 0;
>>  
>>  	trace_ocfs2_file_aio_write(inode, file, file->f_path.dentry,
>>  		(unsigned long long)OCFS2_I(inode)->ip_blkno,
>> @@ -2230,12 +2237,17 @@ static ssize_t ocfs2_file_write_iter(struct kiocb 
> *iocb,
>>  		file->f_path.dentry->d_name.name,
>>  		(unsigned int)from->nr_segs);	/* GRRRRR */
>>  
>> +	if (!direct_io && nowait)
>> +		return -EOPNOTSUPP;
>> +
>>  	if (count == 0)
>>  		return 0;
>>  
>> -	direct_io = iocb->ki_flags & IOCB_DIRECT ? 1 : 0;
>> -
>> -	inode_lock(inode);
>> +	if (direct_io && nowait) {
>> +		if (!inode_trylock(inode))
>> +			return -EAGAIN;
>> +	} else
>> +		inode_lock(inode);
>>  
>>  	/*
>>  	 * Concurrent O_DIRECT writes are allowed with
>> @@ -2244,9 +2256,13 @@ static ssize_t ocfs2_file_write_iter(struct kiocb 
> *iocb,
>>  	 */
>>  	rw_level = (!direct_io || full_coherency || append_write);
>>  
>> -	ret = ocfs2_rw_lock(inode, rw_level);
>> +	if (direct_io && nowait)
>> +		ret = ocfs2_try_rw_lock(inode, rw_level);
>> +	else
>> +		ret = ocfs2_rw_lock(inode, rw_level);
>>  	if (ret < 0) {
>> -		mlog_errno(ret);
>> +		if (ret != -EAGAIN)
>> +			mlog_errno(ret);
>>  		goto out_mutex;
>>  	}
>>  
>> @@ -2260,9 +2276,13 @@ static ssize_t ocfs2_file_write_iter(struct kiocb 
> *iocb,
>>  		 * other nodes to drop their caches.  Buffered I/O
>>  		 * already does this in write_begin().
>>  		 */
>> -		ret = ocfs2_inode_lock(inode, NULL, 1);
>> +		if (nowait)
>> +			ret = ocfs2_try_inode_lock(inode, NULL, 1);
>> +		else
>> +			ret = ocfs2_inode_lock(inode, NULL, 1);
>>  		if (ret < 0) {
>> -			mlog_errno(ret);
>> +			if (ret != -EAGAIN)
>> +				mlog_errno(ret);
>>  			goto out;
>>  		}
>>  
>> @@ -2277,9 +2297,17 @@ static ssize_t ocfs2_file_write_iter(struct kiocb 
> *iocb,
>>  	}
>>  	count = ret;
>>  
>> -	ret = ocfs2_prepare_inode_for_write(file, iocb->ki_pos, count);
>> +	if (direct_io && nowait) {
>> +		if (!ocfs2_overwrite_io(inode, iocb->ki_pos, count, 0)) {
>> +			ret = -EAGAIN;
>> +			goto out;
>> +		}
>> +	}
>> +
>> +	ret = ocfs2_prepare_inode_for_write(file, iocb->ki_pos, count, !nowait);
>>  	if (ret < 0) {
>> -		mlog_errno(ret);
>> +		if (ret != -EAGAIN)
>> +			mlog_errno(ret);
>>  		goto out;
>>  	}
>>  
>> @@ -2355,6 +2383,7 @@ static ssize_t ocfs2_file_read_iter(struct kiocb *iocb,
>>  	int ret = 0, rw_level = -1, lock_level = 0;
>>  	struct file *filp = iocb->ki_filp;
>>  	struct inode *inode = file_inode(filp);
>> +	int nowait = iocb->ki_flags & IOCB_NOWAIT ? 1 : 0;
>>  
>>  	trace_ocfs2_file_aio_read(inode, filp, filp->f_path.dentry,
>>  			(unsigned long long)OCFS2_I(inode)->ip_blkno,
>> @@ -2374,9 +2403,14 @@ static ssize_t ocfs2_file_read_iter(struct kiocb 
> *iocb,
>>  	 * need locks to protect pending reads from racing with truncate.
>>  	 */
>>  	if (iocb->ki_flags & IOCB_DIRECT) {
>> -		ret = ocfs2_rw_lock(inode, 0);
>> +		if (nowait)
>> +			ret = ocfs2_try_rw_lock(inode, 0);
>> +		else
>> +			ret = ocfs2_rw_lock(inode, 0);
>> +
>>  		if (ret < 0) {
>> -			mlog_errno(ret);
>> +			if (ret != -EAGAIN)
>> +				mlog_errno(ret);
>>  			goto bail;
>>  		}
>>  		rw_level = 0;
>> @@ -2393,9 +2427,11 @@ static ssize_t ocfs2_file_read_iter(struct kiocb 
> *iocb,
>>  	 * like i_size. This allows the checks down below
>>  	 * generic_file_aio_read() a chance of actually working.
>>  	 */
>> -	ret = ocfs2_inode_lock_atime(inode, filp->f_path.mnt, &lock_level);
>> +	ret = ocfs2_inode_lock_atime(inode, filp->f_path.mnt, &lock_level,
>> +				     !nowait);
> 
> Should we judge if the flags is included O_DIRECT?
> 
>>  	if (ret < 0) {
>> -		mlog_errno(ret);
>> +		if (ret != -EAGAIN)
>> +			mlog_errno(ret);
>>  		goto bail;
>>  	}
>>  	ocfs2_inode_unlock(inode, lock_level);
>> diff --git a/fs/ocfs2/mmap.c b/fs/ocfs2/mmap.c
>> index 098f5c7..fb9a20e 100644
>> --- a/fs/ocfs2/mmap.c
>> +++ b/fs/ocfs2/mmap.c
>> @@ -184,7 +184,7 @@ int ocfs2_mmap(struct file *file, struct vm_area_struct 
> *vma)
>>  	int ret = 0, lock_level = 0;
>>  
>>  	ret = ocfs2_inode_lock_atime(file_inode(file),
>> -				    file->f_path.mnt, &lock_level);
>> +				    file->f_path.mnt, &lock_level, 1);
>>  	if (ret < 0) {
>>  		mlog_errno(ret);
>>  		goto out;
>> diff --git a/fs/ocfs2/ocfs2_trace.h b/fs/ocfs2/ocfs2_trace.h
>> index a0b5d00..e2a11aa 100644
>> --- a/fs/ocfs2/ocfs2_trace.h
>> +++ b/fs/ocfs2/ocfs2_trace.h
>> @@ -1449,20 +1449,22 @@
>>  
>>  TRACE_EVENT(ocfs2_prepare_inode_for_write,
>>  	TP_PROTO(unsigned long long ino, unsigned long long saved_pos,
>> -		 unsigned long count),
>> -	TP_ARGS(ino, saved_pos, count),
>> +		 unsigned long count, int wait),
>> +	TP_ARGS(ino, saved_pos, count, wait),
>>  	TP_STRUCT__entry(
>>  		__field(unsigned long long, ino)
>>  		__field(unsigned long long, saved_pos)
>>  		__field(unsigned long, count)
>> +		__field(int, wait)
>>  	),
>>  	TP_fast_assign(
>>  		__entry->ino = ino;
>>  		__entry->saved_pos = saved_pos;
>>  		__entry->count = count;
>> +		__entry->wait = wait;
>>  	),
>> -	TP_printk("%llu %llu %lu", __entry->ino,
>> -		  __entry->saved_pos, __entry->count)
>> +	TP_printk("%llu %llu %lu %d", __entry->ino,
>> +		  __entry->saved_pos, __entry->count, __entry->wait)
>>  );
>>  
>>  DEFINE_OCFS2_INT_EVENT(generic_file_aio_read_ret);
>>