|
From: | Max Reitz |
Subject: | Re: [PATCH v4 for-6.0? 0/3] qcow2: fix parallel rewrite and discard (rw-lock) |
Date: | Tue, 30 Mar 2021 11:49:06 +0200 |
User-agent: | Mozilla/5.0 (X11; Linux x86_64; rv:78.0) Gecko/20100101 Thunderbird/78.8.0 |
On 25.03.21 20:12, Vladimir Sementsov-Ogievskiy wrote:
ping. Do we want it for 6.0?
I’d rather wait. I think the conclusion was that guests shouldn’t hit this because they serialize discards?
There’s also something Kevin wrote on IRC a couple of weeks ago, for which I had hoped he’d sent an email but I don’t think he did, so I’ll try to remember and paraphrase as well as I can...
He basically asked whether it wouldn’t be conceptually simpler to take a reference to some cluster in get_cluster_offset() and later release it with a to-be-added put_cluster_offset().
He also noted that reading is problematic, too, because if you read a discarded and reused cluster, this might result in an information leak (some guest application might be able to read data it isn’t allowed to read); that’s why making get_cluster_offset() the point of locking clusters against discarding would be better.
This would probably work with both of your solutions. For the in-memory solutions, you’d take a refcount to an actual cluster; in the CoRwLock solution, you’d take that lock.
What do you think? Max
[Prev in Thread] | Current Thread | [Next in Thread] |