When creating a new transacted storage object (or reverting an
existing one), rather than copy the original storage, we simply create
a "stub directory entry" for the root. As stub entries are accessed,
we fill in their data from the parent and create new stubs for any
linked entries. The streams have copy on write semantics - reads are
from the original entry until a change is made, then we make a copy in
the scratch file.
When committing transacted storages, we have to create a new tree with
the new data so that the storage entry can be modified in one step,
but unmodified sections of the tree can now be shared between the new
tree and the old. An entry can be shared if it and all entries
reachable from it are unmodified. In the trivial case where nothing
has been modified, we don't have to make a new tree at all.
A big block chain is a linked list, and we pretty much need random
access to them. This should theoretically make accessing a random
point in the chain O(log2 n) instead of O(n) (with disk access scaling
based on the size of the read/write, not its location). It
theoretically takes O(n) memory based on the size, but it can do
better if the chain isn't very fragmented (which I believe will
generally be the case for long chains). It also involves fetching all
the big block locations when we open the chain, but we already do that
anyway (and it should be faster to read it all in one go than
piecemeal).
This value is stored in the storage file header. We currently hard-code it to
0x1000. I don't expect to see files in the wild with other values, but
according to MS this is a valid configuration. For now, just fail if we see
another value.
I've also upgraded the message for unexpected values in storage file headers
to a fixme, since they are valid according to MS.
There's no need for this. Any useful information we could get out of it
is availble from the olemalloc channel, and it means that the ole channel
changes behavior in a way that's visible to programs.
We can't determine the correct block size until we read the header, and we
can't create the BigBlockFile until we know the correct block size. To
solve this dilemma, have StorageImpl calculate the file size it needs instead
of asking the BigBlockFile to "ensure a big block exists".
Marcus Meissner