On Mon 19-01-15 22:07:01, Namjae Jeon wrote: > > > When this state is set, any process which tries to modify the file's address > > > space, either by pagefault mmap writes or using write(2), will block until > > > the this state is cleared. I_WRITE_FREEZED is set by calling FS_IOC_FWFREEZE > > > ioctl and clear by FS_IOC_FWTHAW ioctl. > > > > > > File write freeze functionality, when used in conjunction with > > > inode's immutable flag can be used for creating truly stable file snapshots > > > wherein write freeze will prevent any modification to the file from already > > > open file descriptors and immutable flag will prevent any new modification > > > to the file. One of the intended uses for stable file snapshots would be in > > > the defragmentation applications which defrags single file. > > > > I don't quite understand why the full filesystem freeze is > > necessary? The thaw occurs immediately after I_WRITE_FREEZED is set, > We started by looking at fs freeze for file freeze implementation, > So got biased for using fs freeze or similar approach. > Thanks for suggesting a better way. > > > which means there's nothing that prevent the file from being > > truncated or otherwise modified by fallocate, etc while it is > > frozen.... > Right, So, After that, we had also thought of setting immutable > flag of inode. Immutable flag + I_WRITE_FROZEN => truly frozen file. > > > > > AFAICT, fsync will bring the file down to a consistent state and > > we've already got freeze hooks for all inode modification > > operations. We also have IO barriers for truncate operations so that > > we can wait for all outstanding IO to complete, so I would have > > thought this covers all bases for an inode freeze. i.e.: > Right. > > > > > i_mutex -> I_FROZEN -> fsync -> inode_dio_wait > > > > Should give us a clean inode where there are not ongoing operations > > by the time that inode_dio_wait() completes. All new modification > > operations need to check I_FROZEN in addition to the superblock > > freeze checks... > I checked the routines where checks for I_FROZEN would be required. > Most of them are Ok but do_unlinkat() confuses me a little. > vfs_unlink is called under parent inode's i_mutex, so we cannot sleep > keeping parent's i_mutex held. > i.e while freezing file, all file in directory are blocked by parent > i_mutex. Is it ok to release parnets->mutex before checking for I_FROZEN > or there is some idea? So I believe Dave thought that you'd just reuse places we currently use to call sb_start_write() / mnt_want_write(). You'd probably have to come up with a function like path_want_write() (takes struct path as an argument) and which will call mnt_want_write(), sb_start_write(), and do appropriate inode freeze handling. Then you replace all calls to mnt_want_write() with calls to path_want_write()... Possibly you can also provide a trivial wrapper for path_want_write() which takes struct file instead. This should also deal with the locking problems you describe above as mnt_want_write() is always called before taking i_mutex. Honza -- Jan Kara <jack@xxxxxxx> SUSE Labs, CR -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
