On 02/11/2015 11:09 AM, Ivan Shapovalov wrote:
On 2015-02-11 at 09:23 +0100, Edward Shishkin wrote:
On 02/10/2015 09:42 PM, Ivan Shapovalov wrote:
On 2014-12-20 at 21:24 +0100, Edward Shishkin wrote:
This is the promised generalization, which is supposed to work for all
discard
offsets and all discard unit sizes without any restrictions.
Complications in comparison with the previous implementation:
In this general case we need "precise" coordinates, where every
individual byte
can be addressed. All local variables, which represent precise
coordinates are
denoted with "prefixes" (a_len, d_off, p_tailp, etc). Local variables,
which represent
"non-precise" coordinates (they are usually of type reiser4_block_nr)
are denoted
without prefixes (start, len, end, tailp, etc).
Blocks, which contain head and tail paddings are now calculated using the
function size_in_blocks(), which actually is an expression for the
minimal number
of blocks containing the precise extent.
The next trouble is "peculiarity in 0", encountered when calculating the
blocks of
head padding. if discard offset is different from 0, then the first
discard unit of the
partition is partial (its other part doesn't belong to our partition, so
we can not
discard it). We handle this peculiarity by an additional check.
In other bits everything is the same.
Possible optimization: If discard unit sizes are always powers of 2,
then it makes
sense to replace "do_div(offset, unit_size)" with "offset & (unit_size -
1)".
Mount options discard.offset=xxx,discard.unit=yyy are to emulate various
discard unit sizes and offsets on devices _without_ trim support (e.g.
HDDs).
This is only for debugging purposes, don't use it for real SSD devices:
the kernel
retrieves the discard parameters on its own.
This patch is against the patch series of Ivan Shapovalov:
http://marc.info/?l=reiserfs-devel&m=141841865432082&w=2
Current status: not well-tested.
Edward.
Hi,
I've found a bug in our implementation (don't know when it appeared,
maybe it was quite some time ago). I've intended to fix it and send
a patch along with description, but I still can't think of a viable fix.
So: the problem is that check_free_blocks() isn't idempotent, because it
allocates blocks if the whole extent is clean. Therefore, it must not be
called for overlapping ranges. However, in some conditions tail padding
of some extent and head padding of next extent may overlap in terms of
disk blocks (gluing code only catches overlapping erase units).
This will yield a false negative when checking the head padding, so it
does not lead to any data losses (just to inefficiency).
You mean that sometimes we perform unneeded checks?
I see nothing criminal, as we don't exceed announced (2N_e)
number of checks, where N_e is number of extents in the
discard set.
No, I didn't talk about that.
As to fixup: I think that we need to set up the local variable
head_is_known_dirty properly..
Hmm, head_is_known_dirty is an optimization: either known dirty
(in which case we skip checking and cut the head), or unknown
(in which case we do the check).
I'm talking about a different scenario:
- tail padding of an extent is clean
- head padding of the next extent is clean
- these two paddings overlap in terms of disk blocks
In this case, the head padding check will yield false ("dirty") because
part of it has been already allocated for the tail padding, but in fact
it is clean. Thus a false negative: the head will be cut while it can be
padded.
Ah, you suspect non-preciseness (leak of "garbage")?
If tail padding of the current extent overlaps with the head padding of
the next extent, then the end of the current extent and the beginning
of the next extent are in the same erase unit. Otherwise we'll end with
contradiction. Correct? Now note that we'll try to glue such extents
(the current one and its right neighbor).
If this gluing failed (the "area" between the extents is dirty), then we
set head_is_known_dirty = 1, so that head padding of the next extent
won't be checked.
If this gluing successful, then we won't check the head of the next
extent. Because we jump to this next extent and try to glue its right
neighbor.
That is, I still don't see any problems.
Edward.
More generally, our check_free_blocks() is not idempotent. After a
positive result has been returned for a given range [A;B),
check_free_blocks() must not be called for any range [A';B') such that
[A;B) ∩ [A';B') ≠ ∅. Otherwise a false negative can be returned.
Actually, during the last night I've *apparently* came up with a
solution for this :) I'll send a patch once I figure out how to explain
it properly in the comments...
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
