On 10/21/2014 09:00 PM, Ivan Shapovalov wrote:
On Tuesday 21 October 2014 at 20:48:59, Edward Shishkin wrote:
On 10/21/2014 08:11 PM, Ivan Shapovalov wrote:
On Tuesday 21 October 2014 at 20:01:55, Edward Shishkin wrote:
On 10/21/2014 06:42 PM, Ivan Shapovalov wrote:
On Tuesday 21 October 2014 at 18:33:11, Edward Shishkin wrote:
[...]
Ok, it is clear, why we can not fail with -ENOSPC when trying to delete
a file from a full partition, yes?
Sure. Please note that making reiser4_trim_fs's allocations BA_RESERVED will
not hinder this: delete_mutex is there for a reason.
Sure, and this is the reason, because when I hear "let's take a mutex",
I start to feel bad :)
Does this (ab)use of delete_mutex violates any ordering?
If not, it should be OK.
I want to see explanation, why FITRIM ioclt can not finish the work when
there is no free space on disk.
Suppose we do not use reserved space for reiser4_trim_fs's allocations.
Let's analyze those two cases:
1. There is <= 5% free space on disk.
Initial grabbing fails, nothing can be trimmed.
This is wrong.
2. There is 5% + X (where X is some small number) free space on disk.
We can grab only X blocks at a time, so a total of ((SIZE * 5% / X) + 1)
transactions will be created. BTW, if X < erase unit, nothing can be trimmed.
This is ineffective.
Hope this makes sense.
This is already much better, thanks!
Then I should say that the whole idea to reserve % of free space is
incorrect.
Try to reserve 1% of total partition size with BA_CAN_COMMIT for an
iteration.
If it fails, then try to reserve 1 block with BA_CAN_COMMIT.
Hm? Both described cases are still broken.
In #1 (free space <= 5%) nothing changes; reiser4_trim_fs will never be able
to do its task without using reserved space.
In updated #2 (5% < free space < 6%), a transaction per every block (!)
will be created, and none of them will trim anything (1 block < erase unit).
If (5% < free_space < 5% + erase_unit), then fail with ENOSPC.
...and s/1 block/1 erase unit/ in your proposal?
Still this will create too many transactions.
Why not? This is not the unlink case...
It looks like fighting an artificially created problem. I fail to see why
can't we use BA_RESERVED.
Let's assign you a pair of tickets with deadlock issues
in reiser4? I think you will change your opinion quickly :)
This will solve both cases at no harm.
Note -- I'm not saying "let's take this resource, it does not harm", but
"let's take this resource, it solves two cases AND does not harm" :)
OK, you won,
do it with grab_reserved :)
Edward.
--
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
