On 06/30/2014 03:24 PM, Ivan Shapovalov wrote:
On Monday 30 June 2014 at 14:19:33, Edward Shishkin wrote:
Hi Ivan,
I think that everything should be quite simple.
Let the FITRIM ioctl spawn a process, which scans all the bitmap blocks
and puts the free-space-extents to the deferred delete_set. At the same
time, we mark them as "busy" in the WORKING bitmap to make sure that
nobody will touch them (the post_commit_hook will make them clean
again). Basically, that's all: our transaction manager and the discard
code will do all the other work :)
As I've stated, I'm afraid that it will be very inoptimal. The current
algorithm used for "real-time" discard is not suited for discarding
big chunks of known free space:
So who knows the size of free-space extents when we run the ioctl?
it verifies each discard unit by
querying the bitmap. (This is even worse as current implementation
is simplified to do one bitmap query per each discard unit, not checking
if a bitmap block is bigger than a discard unit,
Why do we need to check it? One bitmap block covers 128MiB.
Are there such discard units in the nature?
thus it can result
in doing multiple bitmap queries for the same block.)
However, we can add a per-transaction flag "batch mode discard" which changes
the discard algorithm to something suitable for this scenario.
If this "discard transaction" can be merged with usual ones, we still
need to verify all extents... But it can be done not so thoroughly.
- do one bitmap query per each recorded extent
- if OK, then discard
- if not, just skip the whole extent
The only optimization I can see is skipping the phases of sorting and
merge. Because we scan bitmaps from left to right and nobody
breaks the order of extents in the delete set (assuming that nobody
joins our atom, see below). In other bits everything looks like in the
case of "real-time" discard, that is like "capturing" free-space extents
of random offsets and sizes. No?
Details
First, the trim process reads up a bitmap node and puts it to the
transaction (see try_capture() and friends). Then we go from left to
Hm, bitmap nodes are never captured by current code, so won't this effectively
be a no-op?
Perhaps, my information about the transaction manages is outdated.
We don't fuse atoms on common bitmap nodes? Excellent. Don't capture
bitmaps then and skip the sorting phase. Just add a check that extents
go in the right order when issuing discard requests.
Edward.
right in the region of blocks covered by the bitmap node and handle
the "free extents". In every such iteration we lock the bitmap node,
locate the free extent, mark it dirty in the WORKING bitmap, put the
respective entry to the deferred delete_set and unlock the bitmap node.
This can be done on allocator level by adding an "iterate free space"
interface method...
NOTE that commit of atoms spawned by the trim process will be unusual:
no dirty jnodes, hence, no flush, no IOs (as RELOCATE and OVERWRITE
sets are empty). Only issuing discard requests.. Of course, this is
only in the case when other processes spawning dirty jnodes didn't
join our atom (they can do it perfectly, as we lock bitmaps only per a
free extent handling).
Again, bitmap nodes are not captured by current code. Is it critical that we
fuse with anything modifying the bitmaps?
Thanks,
--
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
