On 06/16/2014 01:00 PM, Ivan Shapovalov wrote:
On Monday 16 June 2014 at 11:24:39, Edward Shishkin wrote:
On 06/16/2014 07:03 AM, Ivan Shapovalov wrote:
On Monday 16 June 2014 at 02:14:49, Edward Shishkin wrote:
On 06/15/2014 11:58 PM, Ivan Shapovalov wrote:
On Sunday 15 June 2014 at 23:49:59, Edward Shishkin wrote:
On 06/15/2014 08:07 PM, Ivan Shapovalov wrote:
On Sunday 15 June 2014 at 19:36:05, Edward Shishkin wrote:
On 06/13/2014 10:28 PM, Ivan Shapovalov wrote:
Here is my "analysis" of what happens in reiser4 during a transaction's
lifetime wrt. block allocation and deallocation.
THE EFFECTS (SEMANTICS) OF RELATED FUNCTIONS
reiser4_alloc_blocks_bitmap(): allocates in WORKING BITMAP
Yes.
reiser4_dealloc_blocks_bitmap(!BA_DEFER): deallocates from WORKING BITMAP
reiser4_dealloc_blocks_bitmap(BA_DEFER): stores to ->delete_set
This is correct for the middle-level allocator (without suffix "bitmap").
The low-level one frees blocks only in WORKING BITMAP.
Yes, this was a wording mistake. I've noticed it shortly after sending...
reiser4_pre_commit_hook_bitmap(): allocates all relocated nodes in COMMIT BITMAP
"Relocated" is bad term here: nodes with new data also get the flag
JNODE_RELOC. So, I would rather say, that it applies freshly allocated
nodes of the atom to COMMIT BITMAP.
deallocates ->delete_set from COMMIT BITMAP
applies the atom's delete_set to COMMIT BITMAP
I've meant exactly this.
reiser4_post_commit_hook(): deallocates ->delete_set using !BA_DEFER
(i. e. from WORKING BITMAP)
applies the atom's delete_set to WORKING BITMAP.
ditto
I would also mention a function of the middle-level block allocator
reiser4_alloc_blocks(): allocates blocks in WORKING BITMAP.
Yes, sure.
Note that the middle-level block allocator (block_alloc.c) actually
manipulates
with abstract space maps. Currently in reiser4 they are represented only by
bitmaps (plugin/space/bitmap.c). We can also implement another
representation -
extent tree (like in XFS). I don't see any needs for now, though.
Extent trees seem to be interesting.. They claim that they are more efficient -
is the difference that huge?
TIMELINE OF ALLOCATIONS FOR "USUAL" NODES, AND TIMELINE OF TRANSACTION COMMIT
- nodes are allocated using reiser4_alloc_blocks() and setting JNODE_RELOC,
so WORKING BITMAP ensures that two nodes cannot get the same block;
- nodes are deallocated using reiser4_dealloc_blocks(BA_DEFER),
so their deallocation is not immediately reflected in WORKING BITMAP;
(the relocate set is written here)
- reiser4_pre_commit_hook_bitmap() uses 1) JNODE_RELOC flag and 2) ->delete_set
to convey effective bitmap changes into COMMIT BITMAP;
(the journal and overwrite set are written here)
- reiser4_post_commit_hook() uses ->delete_set to convey deallocations
from step 2 to WORKING BITMAP.
(the discard happens here)
TIMELINE OF ALLOCATIONS FOR WANDERED JOURNAL BLOCKS
- at commit time, blocks are allocated using reiser4_alloc_blocks(), so they
are allocated in WORKING BITMAP and do not interfere with any "usual" blocks;
- after writing wandered blocks, they are deallocated using
reiser4_dealloc_blocks(!BA_DEFER), i. e. from the WORKING BITMAP.
So, the system of working and commit bitmaps plus the delete set seems
to be redundant? I think this is because of performance reasons: block
allocation is critical thing...
Seems like it is -- for performance and simplicity (== robustness).
CONCLUSION
At possible transaction replay time, journal blocks are not allocated in any
of the bitmaps. However, because the journal is read and replayed before a
transaction has a chance to commit, this fact does not matter.
What matters is that wandered journal blocks never hit COMMIT BITMAP.
So, if I've got all this correct (which I highly doubt), the disk space leak
(as you pointed it out) does not exist.
It seems, you are right..
What exists is a rather different problem with my idea of "log every
deallocated block". Current implementation logs every block regardless of
BA_DEFER flag presence or absence, so non-wandered blocks are logged twice.
We could just use ->delete_set, but we would lose wandered blocks then.
Or we could only log !BA_DEFER requests, which would do the right thing
(wandered blocks + deallocations from reiser4_post_commit_hook()), but
the reasoning behind this decision would not be obvious for a casual
code reader.
I think that a good comment will save the situation..
Or we could log only wandered blocks (in addition to ->delete_set)
at discard time, but this is messy and requires us to merge the discard log
with ->delete_set at discard time.
what is the difference with the previous "we could.."?
The previous option is to log all !BA_DEFER requests in addition to
->delete_set, so those two block sets would be partially overlapping.
Hmm.. Are they really overlapped?
reiser4_dealloc_blocks() looks like the following:
{
if (flags & BA_DEFER)
log in the delete_set;
else
deallocate in working bitmap
...
}
so (!BA_DEFER) requests and the delete_set look disjoint,
or, I missed something?
You're right, I've misread the code. I was under impression that
reiser4_post_commit_hook() internally calls reiser4_dealloc_blocks(!BA_DEFER).
So the second option (which I prefer) is to log all sa_dealloc_blocks() calls.
That is, effectively, ->delete_set plus wandered journal block deallocations.
Yes, I think to do something like this:
reiser4_dealloc_blocks()
{
if (flags & BA_DEFER) {
...
} else {
...
if (discard is turned on)
/* we'll want to discard these blocks, which are not to
be represented
in the COMMIT SPACE MAP, so store them in a
separate list */
do {
blocknr_set_add_extent(atom,
&atom->delete_set_for_wander, &bsep, start, len);
...
} while (ret == -E_REPEAT);
sa_dealloc_blocks(reiser4_get_space_allocator(ctx->super), *start, *len);
...
}
Then join the sets (->delete_set, and ->delete_set_for_wander) at
discard time
(should be pretty cheap operation). insert a check at every atoms
fusion, that
->delete_set_for_wander of both atoms are empty).
Hm. IIRC, ultimately we wanted to use rb-trees for storing discard block sets,
but this means that we'll need to convert two blocknr_sets into an rb-tree
at discard time, which pretty defeats their purpose in our context... Or do you
say that blocknr_set itself should be made rb-tree?
I didn't know about the delete sets. They are not optional and lists are
better for them (as they don't require sorting). So I suggest to confine
with lists for now, and may be to try the fast-merge trees in the future
(if there is a great desire).
The discard algorithm still needs its input to be sorted, and blocknr_set is
inherently unordered.
So there are two options:
- log into ->delete_set and ->delete_set_for_wander (as you've proposed),
then merge them into a sorted list at discard time;
- log into ->discard_set (just like it is now), just fix what to log.
What do you think?
Are the blocknr sets comfortable?
Say, can we organize the discard process via the blocknr_set_iterator()?
If yes, then let's do everything via the blocknr sets (i.e. let's
implement the
first option). Add needed operations to blocknrset.c, the discard iteration
actor to discard.c, etc.
--
To unsubscribe from this list: send the line "unsubscribe reiserfs-devel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
