Re: HAMMER update 23-Mar-08

[Matthew Dillon <dillon@xxxxxxxxxxxxxxxxxxxx>]

:Wouldn't making timestamp queries (at least from userland) enforce a 
:sync on the volume in question be useful here?
:-- 
:         Thomas E. Spanjaard
:         tgen@netphreax.net

    Making the 'hammer now' command do a sync() is a good idea.  I will
    make that change right now so it doesn't get lost.

    Here's a general overview of the issues involved with having historical
    access to the filesystem:

				----------------

    Recording the timestamps in the in-memory cache, for a finer-grained
    snapshot capability, is doable but has its own issues.  Here's an
    illustration:

    open()  create file
    write() append 4K	(file size now 4K)
    write() append 4K	(file size now 8K)
    write() append 4K	(file size now 12K)
    write() append 4K	(file size now 16K)

    Now NONE of this has gone to disk yet, it's entirely in the in-memory
    cache.  The inode is in the in-memory cache.  The data is stored in
    the buffer cache.  Even the directory entry for the file that we
    just created is still in the in-memory cache (HAMMER caches the raw
    records it intends to commit later on).

    If I wanted to be able to acquire a timestamp between each write and
    'see' a snapshot of the file as of any point in the above sequence,
    then every write would also have to allocate a copy of the inode
    (because it changes size on each write).

    The data has the same problem though with a slightly different example.
    Lets say each write() was a seek-write, overwriting the previous data.
    Now with every write() I would have to allocate a copy of the data
    being overwritten.  This is complicated by the fact that the buffer
    cache has no clue about 'historical' accesses, so I would not be able
    to use the buffer cache to cache the data.

    There's also another problem and that is with the efficiency of the
    topology on-disk.  Even if I maintained all the copies of the inode
    and all the copies of the data in-memory, I would still have to sync
    all those copies to disk in order for things to remain historically
    coherent (whether it be in-cache or on-disk).  This would result in
    hundreds or even thousands of copies of the inode on-disk, not to
    mention potentially many copies of the data.

    I just don't want to do that right now, at least not as a default.  A
    lot of performance would be lost.  Hence a sync() is needed if
    you want to create a demark which you can accurately snapshot.

				-------------

    Here's a quick synopsis of how the cache would operate in a clustered
    filesystem:

    In order to properly integrate with in-memory caches, a wider cache
    coherency infrastructure is needed between machines such that
    modifications made on one machine proactively invalidate those
    protions of the cache(s) on other machines.  At the same time, any
    'dirty' cache data, for example when a file is created or written to,
    must lock the cache space in question on all other machines.
    The cache space in this case is not just the file data, but also the
    related namespaces (for creations, deletions, and renames).  Attempts
    to access locked spaces from other machines in the cluster would have
    to force a flush to the filesystem backing store and lower the cache
    states for the effected information on the original machine from dirty
    to shared-read-only.

    It will be easiest to integrate the cache coherency information into
    the buffer cache and namecache themselves.  Once a machine has dirtied
    an in-memory cache element... for example part of the namespace when
    creating a file or chunks of data written within a file, that machine
    must have a free hand to make further modifications to the cache spaces
    involved without further interaction with other machines.

				-------------

    Now, if you think of those two major elements you can see that they
    actually fit together quite well.  If I were to attempt to maintain
    transactional coherency on a per-system-call basis then the cache
    granularity between machines would have to be much, much smaller then
    our current in-memory caching elements provide.  That would become
    a really nasty coding problem.  So I don't even want to begin to
    complate transactional coherency at a finer-grain then sync() or
    fsync() until long after we actually have clustering working.

						-Matt