12.03.2021 18:52, Max Reitz wrote:
On 12.03.21 16:24, Vladimir Sementsov-Ogievskiy wrote:
12.03.2021 18:10, Max Reitz wrote:
On 12.03.21 13:46, Vladimir Sementsov-Ogievskiy wrote:
12.03.2021 15:32, Vladimir Sementsov-Ogievskiy wrote:
12.03.2021 14:17, Max Reitz wrote:
On 12.03.21 10:09, Vladimir Sementsov-Ogievskiy wrote:
11.03.2021 22:58, Max Reitz wrote:
On 05.03.21 18:35, Vladimir Sementsov-Ogievskiy wrote:
There is a bug in qcow2: host cluster can be discarded (refcount
becomes 0) and reused during data write. In this case data write may
[..]
@@ -885,6 +1019,13 @@ static int QEMU_WARN_UNUSED_RESULT
update_refcount(BlockDriverState *bs,
if (refcount == 0) {
void *table;
+ Qcow2InFlightRefcount *infl = find_infl_wr(s, cluster_index);
+
+ if (infl) {
+ infl->refcount_zero = true;
+ infl->type = type;
+ continue;
+ }
I don’t understand what this is supposed to do exactly. It seems like it wants
to keep metadata structures in the cache that are still in use (because
dropping them from the caches is what happens next), but users of metadata
structures won’t set in-flight counters for those metadata structures, will
they?
Don't follow.
We want the code in "if (refcount == 0)" to be triggered only when full reference count
of the host cluster becomes 0, including inflight-write-cnt. So, if at this point
inflight-write-cnt is not 0, we postpone freeing the host cluster, it will be done later from
"slow path" in update_inflight_write_cnt().
But the code under “if (refcount == 0)” doesn’t free anything, does it? All I
can see is code to remove metadata structures from the metadata caches (if the
discarded cluster was an L2 table or a refblock), and finally the discard on
the underlying file. I don’t see how that protocol-level discard has anything
to do with our problem, though.
Hmm. Still, if we do this discard, and then our in-flight write, we'll have
data instead of a hole. Not a big deal, but seems better to postpone discard.
On the other hand, clearing caches is OK, as its related only to
qcow2-refcount, not to inflight-write-cnt
As far as I understand, the freeing happens immediately above the “if (refcount ==
0)” block by s->set_refcount() setting the refcount to 0. (including updating
s->free_cluster_index if the refcount is 0).
Hmm.. And that (setting s->free_cluster_index) what I should actually prevent
until total reference count becomes zero.
And about s->set_refcount(): it only update a refcount itself, and don't free
anything.
So, it is more correct like this:
diff --git a/block/qcow2-refcount.c b/block/qcow2-refcount.c
index 464d133368..1da282446d 100644
--- a/block/qcow2-refcount.c
+++ b/block/qcow2-refcount.c
@@ -1012,21 +1012,12 @@ static int QEMU_WARN_UNUSED_RESULT
update_refcount(BlockDriverState *bs,
} else {
refcount += addend;
}
- if (refcount == 0 && cluster_index < s->free_cluster_index) {
- s->free_cluster_index = cluster_index;
- }
s->set_refcount(refcount_block, block_index, refcount);
if (refcount == 0) {
void *table;
Qcow2InFlightRefcount *infl = find_infl_wr(s, cluster_index);
- if (infl) {
- infl->refcount_zero = true;
- infl->type = type;
- continue;
- }
-
table = qcow2_cache_is_table_offset(s->refcount_block_cache,
offset);
if (table != NULL) {
@@ -1040,6 +1031,16 @@ static int QEMU_WARN_UNUSED_RESULT
update_refcount(BlockDriverState *bs,
qcow2_cache_discard(s->l2_table_cache, table);
}
+ if (infl) {
+ infl->refcount_zero = true;
+ infl->type = type;
+ continue;
+ }
+
+ if (cluster_index < s->free_cluster_index) {
+ s->free_cluster_index = cluster_index;
+ }
+
if (s->discard_passthrough[type]) {
update_refcount_discard(bs, cluster_offset, s->cluster_size);
}
I don’t think I like using s->free_cluster_index as a protection against
allocating something before it.
Hmm, I just propose not to update it, if refcount reached 0 but we still have
inflight writes.
First, it comes back the problem I just described in my mail from 15:58 GMT+1,
which is that you’re changing the definition of what a free cluster is. With
this proposal, you’re proposing yet a new definition: A free cluster is
anything with refcount == 0 after free_cluster_index.
I think that free cluster is anything with refcount = 0 and inflight-write-cnt
= 0.
Then, as I said in my other mail, update_refcount() just cannot free any
cluster. So changes to that function can’t be justified by preventing it from
freeing clusters.
You need to clearly define what it is that update_refcount() should or
shouldn’t do, and then we have to think about whether when all writes have
settled, we really have to invoke qcow2_update_cluster_refcount() or whether we
should do the small outstanding changes just directly in
update_inflight_write_cnt().
I think this needs to be more formalized, or it doesn’t make sense.
For example, say we do define a free cluster to be refcount (RC) = 0 and
inflight-write-cnt (IFWC) = 0. Then everything that is done to a cluster because
it is considered being freed right now because its RC drops to 0 must probably be
changed to only be done if also its IFWC is 0. For example, we should only
discard host clusters on the protocol layer if a cluster becomes free.
update_refcount() will no longer be able to free clusters with IFWC > 0, so it
must never issue a protocol-level discard for them. And, yes, it also shouldn’t
adjust first_free_cluster_index, as you propose here. (But you didn’t explain
why, and it seems like it was just intuition to you instead of looking at it more
formally.)
Instead, for clusters with RC = 0 and IFWC > 0, update_inflight_write_cnt()
will take on the role of freeing them. So now that function must adjust
first_free_cluster_index and issue the protocol-level discard for such clusters.
Yes, agree.
I suppose in practice we could invoke qcow2_update_cluster_refcount() with -0,
as you do, because now the cluster has RC = 0 and IFWC = 0, so now that
function will be capable of freeing it. But to me, that just looks like a bit
of abuse.
agree
I suppose we could create a new function qcow2_cluster_freed() where we collect
everything that needs to be done once a cluster is considered freed (which so
far was whenever its RC dropped to 0, which only happens in update_refcount();
and then will be whenever its RC and its IFWC drop to 0, which can happen in
either update_refcount() or update_inflight_write_cnt()). What would belong in
there is discarding the cluster on the protocol level, and adjusting
first_free_cluster_index. (Perhaps more, I don’t know.) With such a function,
it would seem clear to me that there is no need to invoke
qcow2_update_cluster_refcount() just to get precisely that effect.
yes
(The alternative would be to keep RC == 0 the definition of a freed cluster. Then
we’d have to postpone the s->set_refcount() in update_refcount(), and update
the refcount again in update_inflight_write_cnt(), but invoking
qcow2_update_cluster_refcount(). We wouldn’t need to change the allocation
functions.
I’m not saying that alternative is better – I don’t think it is, I think you’re
right that the definition of a freed cluster should be changed. I’m just
presenting it in contrast, to show when it would make sense to call
qcow2_update_cluster_refcount().)
OK
In the meanwhile Kevin dispelled my "big problems" in "[PATCH v2(RFC) 0/3] qcow2:
fix parallel rewrite and discard", so probably next step would be to retry CoRwLock-based
approach.
And free_cluster_index is a hint where start to search for such cluster.
Now looking only at the allocation functions, it may look like that kind of is
the definition already. But I don’t think that was the intention when
free_cluster_index was introduced, so we’d have to check every place that sets
free_cluster_index, to see whether it adheres to this definition.
And I think it’s clear that there is a place that won’t adhere to this
definition, and that is this very place here, in update_refcount(). Say
free_cluster_index is 42. Then you free cluster 39, but there is a write to
it, so free_cluster_index isn’t update. Then you free cluster 38, and there
are writes to that cluster, so free_cluster_index is updated to 38. Suddenly,
39 is free to be allocated, too.
Why? 39 is protected by inflight-cnt, and we do has_infl_wr() check together
with refcount==0 check when allocate clusters.
I was (wrongly) assuming that with this change you’d drop the check in the
allocation functions.
Max
(The precise problem is that with this new definition decreasing
free_cluster_index suddenly has the power to free any cluster between its new
and all value. With the old definition, changing free_cluster_index would
never free any cluster. So when you decrease free_cluster_index, you suddenly
have to be sure that all clusters between the new and old value that have
refcount 0 are indeed to be considered free.)
Max
--
Best regards,
Vladimir