TODO

SnapRAID TODO
=============

This is the list of TODO items for SnapRAID.

- Next

* Add an option to ignore a specific Smart attribute. This solves PullRequest #17 an #21

* Support BTRFS subvolumes. Keep track of the subvolumeid, along with the inode, and use it when searching/comparing inodes.
The value 0 can be used for the root volume, like a normal filesystem.
See: https://github.com/kdave/btrfs-progs/tree/master/libbtrfsutil

* A "dependents" output option to [--test-fmt file|disk|path|dependents] for LIST. 
Example, snapraid LIST --test-fmt dependents >FileList_with_dependents.txt
The goal is a filelist that, for every file, includes full paths to the dependent files 
on other data disks needed to repair that file (in addition to the parity data).

I created a thread about it here: https://sourceforge.net/p/snapraid/discussion/1677233/thread/6e872fdae0/
- Minor

* Add a new -u, --filter-updated command that filters files
with a different timestamp, to be able to restore only them to the previous state.
See: https://sourceforge.net/p/snapraid/discussion/1677233/thread/e26e787d/

* Allow to filter in diff/list by disk name.
Not checked disks should be allowed to be missing.

* Add an option to ignore subsecond timestamp.
Like when you copy data to a filesystem with less timestamp precision.

* Use threads to scan all the disks at the same time.
- After 7.0 Windows changes it seems fast enough even
with a mono thread implementation with 100.0000 files.
+ But if you have millions of files, it could take minutes.

* Support more parity levels
It can be done with a generic computation function, using
intrinsic for SSSE3 and AVX instructions.
It would be interesting to compare performance with the hand-written
assembler functions. Eventually we can convert them to use intrinsic also.
https://sourceforge.net/p/snapraid/discussion/1677233/thread/9dbd7581/

* Extend haspdeep to support the SnapRAID hash :
https://github.com/jessek/hashdeep/
https://sourceforge.net/p/snapraid/discussion/1677233/thread/90b0e9b2/?limit=25

* Add a "noaccessdenied" config option to exclude not readable files/dirs.
Like the "nohidden".
See: https://sourceforge.net/p/snapraid/discussion/1677233/thread/1409c26a/

* Don't insert new files if they are new opened by other applications.
Not yet sure how to check if a file is open in a fast way.
In case we can exclude files created too recently, like with a --min-age option.
See: https://sourceforge.net/p/snapraid/discussion/1677233/thread/a1683dd9/?limit=25#1e16

* Add markdown support for output
See: https://sourceforge.net/p/snapraid/discussion/1677233/thread/661cce8b/

* Add ZFS support for SMART and UUID.
See latest messages at: https://sourceforge.net/p/snapraid/discussion/1677233/thread/42defa3b/

* Change all the path printed in the terminal as relative and not absolute.
https://sourceforge.net/p/snapraid/discussion/1677233/thread/648955ec/ (Leifi in diff)
https://sourceforge.net/p/snapraid/discussion/1677233/thread/5b6ef1b9/ (Miles in check + filter)
Partially done. Now you can control it with --test-fmt

* Add ReFS support with 128 bits inode
See: https://sourceforge.net/p/snapraid/discussion/1677233/thread/2be14f63/
https://github.com/Microsoft/Windows-driver-samples/blob/master/filesys/miniFilter/avscan/filter/utility.h
https://github.com/Microsoft/Windows-driver-samples/blob/master/filesys/miniFilter/avscan/filter/utility.c
https://github.com/Microsoft/Windows-driver-samples/blob/master/filesys/miniFilter/delete/delete.c
Partially done. Until the inodes are less than 64 bit, everything works.
If a 128 bit inode is found, it aborts with "Invalid inode number! Is this ReFS?"

* When a disk is replaced, and SnapRAID detect this by a UUID change, we could clear the
scrub information for that disk.
https://sourceforge.net/p/snapraid/discussion/1677233/thread/ee87901b/

* In fix, automatic --import excluded directory in the config file.
https://sourceforge.net/p/snapraid/discussion/1677233/thread/c80c42e5/?limit=25

* Allow to have "partial" parity disks, smaller than required ?
https://sourceforge.net/p/snapraid/discussion/1677233/thread/e924c663/

* How to handle corrupt copied/moved files ? At now pre-hash checks them,
but if a corruption is found, it stops the sync process.
This is not optimal, because the 'sync' command is stopped instead to continue.
A possible solution could be:
If a copied block is corrupt, we can recover the original one from parity (if moved),
or read it directly from disk (if copied), and we can use this one to build the parity,
working around the corruption, that can be later fixed with "fix".
This requires to allocate blocks of moved files in parity positions *after* the
currently used one.

* Some programs are misusing the FILE_ATTRIBUTE_TEMPORARY.
In fact, it makes sense as it's a good way to control the cache on the file,
so, despite the name, that usage could be common for not temporary files.
https://sourceforge.net/p/snapraid/discussion/1677233/thread/57d40108/

* Restore ownership and permissions, at least in Unix.

* Restore directory timestamps.

* Add an option for dup to list only duplicates with different name.
This supposing that if a file has the same name, it's intentionally duplicated.

* In fix an existing symlink with the same name of a file to be recovered may stop
the process making the create() operation to fail.
The same for directories, when recreating the directory tree.

* If a directory exists with the same name of a parity/content file be more explicative
on the error message. See: https://sourceforge.net/projects/snapraid/forums/forum/1677233/topic/4861034

* We don't try to do partial block recovering. A block is correct or not.
But if only some bytes, or a sector, is wrong, it should be possible to recover all the
rest of the block.
The problem is that we don't have any hash to ensure that the block is partially recovered,
or completely garbage. But it makes sense to anyway write the "most likely" correct one.

- Naming

* A new 'init' command to differentiate the first 'sync' operation.
This 'init' will work also without a content file, and parity files.
Instead 'sync' will require all of them.
This will also help when running with the parity filesystem unmounted.

* Rename sync->backup and fix->restore. It seems to me a naming expressing
better the meaning of the commands. But not yet sure.

- Pooling

* Add a new "commit" command to move changes from the pool to the array.
It should:
- Move files copied into the pool (that are no links) to the array.
The files should be moved to the disk that contains most of the files
in the directory. If no space, try with the disk with less files
in the directory, and eventually the disk in the array with more free space.
- Detect renames, and apply them in the array.
The file will be renamed and moved to the new directory, if changed,
but kept in the same disk of the array.
- Detect deletes, and move file in the array to a "/trash/" directory
of the same disk. For safety no real deletion is done.
File with the same name will get an extra extension like ".1", ".2".

- Major

* Uses a separated file for storing hashes, to allow to use a memory
mapped file to decrease memory utilization.
The file can contains hashes stored in the same order they are accessed
in sync/check/scrub.
+ A lot less memory utilization
- It will be slow in --pre-hash as the files are accessed in another order,
but no much slow.
- How to handle multiple .content file copy ? When working we can have
only a single file. When storing the .content we'll have to copy it
in all the places.
* Can we mix the existing and new approach ? We can create this
hash file at startup in a memory mapped "temporary" file.
It may takes some time to create it, but then it will be fast.
See: https://sourceforge.net/p/snapraid/discussion/1677233/thread/cdea773f/

* Allocate parity minimizing concurrent use of it
Each parity allocation should check for a parity
range with less utilization by other disks.
We need to take care do disable this mechanism when the parity space
is near to fillup the parity partition.
See: https://sourceforge.net/p/snapraid/discussion/1677233/thread/1797bf7d/
+ This increase the possibility of recovering with multiple failures with not
enough parity, as the benefit of existing parity is maximized.
- Possible increase of fragmentation of parity allocation
- No real benefit when the array is filled


Rejected TODO
=============

This is a list of rejected TODO items.

* Allow multiple parity files and to coexist with data
in the same disk.
This is possible if the data in the same disk uses parity addresses not
contained in the parity file present in the same disk.
+ It's a nice trick. The disk containing parity, will have less space available,
and then it will need less parity, resolving the problem of the parity disk being too small.
+ We can also think at an automated parity files naming and creation, removing
the need of the user to specify that. We can also completely remove the
concept of parity drive, automatically allocating parity in the most free data drives.
- It won't be not nice to have the program to automatically
choose where to create the parity, because the choice cannot be optimal.
With a dedicated disk manually chosen, it's instead optimal.
+ We can limit this coexist possibility only to latest parity file,
allowing the user to choose where to put it.
- It won't work with disks of different size.
Suppose to have all disks of size N, with only one of size M>N.
To fully use the M space, you can allocate a full N parity in such disk,
but the remaining space will also need additional parity in the other disks,
in fact requiring a total of M parity for the array.
In the end, we cannot avoid that the first biggest disk added is fully
dedicated to parity, even if it means to leave some space unused.

* Directly access disks for parity skipping the filesystem layer
- No real performance or space gain compared to a custom filesystem configuration like:
mkfs.ext3 -i 16777216 -m0 -O ^dir_index,large_file,sparse_super /dev/sdX1
See: https://sourceforge.net/p/snapraid/discussion/1677233/thread/9c0ef324/?limit=25

* Allow to specify more than one disk directories to cover the case of multi partitions.
Different partitions have duplicate inode. The only way to support this is to
add also a kind of device_id, increasing the memory required.
But it should be only few bits for each file. So, it should be manageable.
A lot of discussions about this feature :)
https://sourceforge.net/p/snapraid/discussion/1677233/thread/b2cd9385/
- The only benefit is to distribute better the data. This could help the recovery process,
in case of multiple failures. But no real usability or functionality benefit in the normal
case.

* https://sourceforge.net/p/snapraid/discussion/1677233/thread/2cb97e8a/
Have two hashes for each file. One for the first segment, that may use only a part of the first parity block.
And a second hash that takes care of the final segment, also using only part of a parity block.
Intermediate blocks can be handled like now.
The first and last segment can use only a part of the parity block, sharing it with other files,
and then allowing big parity blocks.
+ Big parity block, like 1MB, and small file allocation, like 4KB.
- It won't be possible to copy hash info from one disk to another, as copied files
may have a different block splitting. Copied file should be allocated with the same exact
splitting.
- Dup detection can be handled with a dedicated hash covering the full file. But increasing
the number of hash for each file to three.
- No way to handle block import.

* Create a filesystem snapshot at every sync, and use it in all the other commands
automatically.
At the next sync, drop the old snapshot and create a new one.
This should help recovering, because we'll have the exact copy used by sync.
This feature can be enabled with a specific option, and available
in Windows using Shadow Copy, and in Linux using Btrfs, and in a generic
Unix using ZFS.
See Jens's Windows script at: http://sourceforge.net/p/snapraid/discussion/1677233/thread/a1707211/
Note that a different between Windows and Unix is that in Windows old snapshots
are automatically deleted.

* Checks if splitting hash/parity computation in 4K pages
can improve speed in sync. That should increase cache locality,
because we read the data two times for hash and parity,
and if we manage to keep it in the cache, we should save time.
- We now hash first the faster disks, and this could
reduce performance as we'll have to wait for all disks.

* Enable storing of creation time NTFS, crtime/birth time EXT4.
But see: http://unix.stackexchange.com/questions/50177/birth-is-empty-on-ext4
coreutils stat has an example, but it doesn't work in Linux (see lib/stat-time.h)
- Not supported in Linux.

* In the content file save the timestamp of the parity files.
If they do not match, stop the processing.
This can be done to avoid to use not synchronized parity and content files,
resulting in wrong data.
But if the sync process is killed we need a way to resyncronize them.
Or maybe we should allow parity newer than content, but not viceversa.
- The corner cases are too many. A fixed parity may be never.
A someway modified content may be never. So, the time is not really significant.

* Use Async IO for Linux (libaio).
See thread: https://sourceforge.net/p/snapraid/discussion/1677233/thread/a300a10b/
Docs at: http://www.fsl.cs.sunysb.edu/~vass/linux-aio.txt
- Implemented for scrub in the "libaio" branch, but it's slower of
about 20% in my machine.

* Allow to put parity directly into the underline block device without the need
of any filesystem.
That would allow to increase a lot the free space for parity.
We can implement some kind of filesystem detection to avoid to overwrite an already
existing filesystem.
- Still risky if stuffs go wrong.
- In Linux with largefile4 there is only a very small amount of space wasted.
In the order of 0.01%. Not really worth to do it.
- With NTFS the saving is also limited, because the "reserved" MFT of 12.5% is
not really exclusively reserved, but can be used also for normal files,
when all the remaining space is filled.

* The data could be compressed before processing, resulting in parity block of
fixed size, but matching different data block sizes.
The complexity is that a file blocks will have to be allocated at runtime,
and you may run out of them in the middle of the processing.
We need also a way to compress a stream until the compressed data reach the
block size, but no more, and then start a new block.
For each block, we'll have also to store. "size_uncompressed", "size_compressed",
"hash".
- It will be too slow.
- Not addressing the problem of a lot of small files, as still one block will be
allocated for each file.

* Uses different block size for parity and file allocation.
We can use a big size, like 1MB for parity allocation and hash computation,
and at the same time using a 4K blocks for files allocation.
This means that a parity blocks may contain more than one file.
- We'll have to drop the dup and import feature,
because it won't be possible anymore to compare the hash of files,
as it would depend also on the start address inside the parity block.
- When a hash fail, it won't be possible to tell which file is really
broken, because more file may share the same parity block.

* Have special parity blocks containing the last blocks of more files
until it's filled up.
For such parity blocks, we'll have more than one hash. One for each
file contained.
- Parity will be too fragmented, because we'll have parity blocks
containing the last blocks of many files.

* In "pool" for Windows, and for unique directories a junction to the
directory could be used, avoiding to use symlinks to files.
This allows to overcome the problem of sharing symlinks.
- It would work, in fact it would work to well. The problem is that
Windows will treat the junction as the real directory, like *really*
deleting its content from Explorer pressing "del" and also from the
command line with "rd". Too dangerous.

* When fixing, before overwriting the present file, make a copy of it just in
case that the original file cannot be completely recovered.
We can always open files in read-only mode, if a write is required, we close it,
rename it to with a .bak extension, and rewrite it up to the required size.
The same for symlink if a file with the same name exist or viceversa.
- The idea behind this is to avoid to leave untouched a file if we cannot
restore it completely. But it's debatable what's better in this case.
Anyway, considering the typical use, it's not really relevant.

* In the import list, uses also all the blocks in the array.
But we must cover the case of bad blocks. Likely we can just check the
hash after reading, and in case, skip it, and retry with another copy.
- It will work only for duplicate files. Not really worth do to it.

* Save the content file in compressed .gz format to save space.
- Compression is too slow. Even using the very fast lzo.
$ time lzop  -1 < content > content.lzo
real    1m23.014s
user    0m40.822s
sys     0m3.389s

$ time ./gzip -1 < content > content.gz
real    1m47.463s
user    1m23.732s
sys     0m3.290s

$ time ./gzip --huffonly < content > contentH.gz
real    1m51.607s
user    1m30.032s
sys     0m3.245s

Similar command done with snapraid without compression, and involving also decoding
and encoding takes less time.

$ time ./snapraid --test-skip-device --test-skip-self -v -c ./test.conf test-rewrite
real    0m59.087s
user    0m14.164s
sys     0m4.398s

* Recognizes that a file is moved from one disk to another, and if the parity
data doesn't overlap, do not recompute it.
- It's going to work only in RAID5 mode and only in special cases.

* Implements a multithread sync command to share HASH and RAID computations
to different CPUs.
- At now it's questionable if it will result in a performance improvement.
The murmur3 hash, and the RAID5/6 computations are so fast that even a single
thread should be able to do them.
Use the "snapraid -T" comment to see the speed.

* In the repair() function the heuristic to detect if we recovered after the sync,
can be extended to all the previous blocks, because we always proceed in block
order during a sync.
So, if for a block we can detect that we recovered using updated parity data,
also for all the previous blocks this is true.
Anyway, the case where this information could be useful should be present
only if changes are committed after an aborted sync.
- No real advantage on that, beside some speed gain in fix.
The risk would be instead to miss some recovery opportunity.
So, makes sense to have it a little slower but trying any
possible recovery strategy.