[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Qemu-block] [Qemu-devel] [PATCH 12/17] migration: add postcopy migr

From: Vladimir Sementsov-Ogievskiy
Subject: Re: [Qemu-block] [Qemu-devel] [PATCH 12/17] migration: add postcopy migration of dirty bitmaps
Date: Mon, 10 Jul 2017 12:17:17 +0300
User-agent: Mozilla/5.0 (X11; Linux x86_64; rv:52.0) Gecko/20100101 Thunderbird/52.2.1

08.07.2017 02:32, John Snow wrote:

On 07/07/2017 05:13 AM, Vladimir Sementsov-Ogievskiy wrote:
06.07.2017 20:53, John Snow wrote:
On 07/06/2017 04:05 AM, Vladimir Sementsov-Ogievskiy wrote:
06.07.2017 00:46, John Snow wrote:
On 07/05/2017 05:24 AM, Vladimir Sementsov-Ogievskiy wrote:
16.02.2017 16:04, Fam Zheng wrote:
+            dbms->node_name = bdrv_get_node_name(bs);
+            if (!dbms->node_name || dbms->node_name[0] == '\0') {
+                dbms->node_name = bdrv_get_device_name(bs);
+            }
+            dbms->bitmap = bitmap;
What protects the case that the bitmap is released before migration

What is the source of such deletion? qmp command? Theoretically

I see the following variants:

1. additional variable BdrvDirtyBItmap.migration, which forbids bitmap

2. make bitmap anonymous (bdrv_dirty_bitmap_make_anon) - it will
not be
available through qmp

Making the bitmap anonymous would forbid us to query the bitmap, which
there is no general reason to do, excepting the idea that a third party
attempting to use the bitmap during a migration is probably a bad idea.
I don't really like the idea of "hiding" information from the user,
though, because then we'd have to worry about name collisions when we
de-anonymized the bitmap again. That's not so palatable.

what do you think?

The modes for bitmaps are getting messy.

As a reminder, the officially exposed "modes" of a bitmap are

FROZEN: Cannot be reset/deleted. Implication is that the bitmap is
otherwise "ACTIVE."
DISABLED: Not recording any writes (by choice.)
ACTIVE: Actively recording writes.

These are documented in the public API as possibilities for
DirtyBitmapStatus in block-core.json. We didn't add a new condition for
"readonly" either, which I think is actually required:

READONLY: Not recording any writes (by necessity.)

Your new use case here sounds like Frozen to me, but it simply does not
have an anonymous successor to force it to be recognized as
"frozen." We
can add a `bool protected` or `bool frozen` field to force recognition
of this status and adjust the json documentation accordingly.
Bitmaps are selected for migration when source is running, so we should
protect them (from deletion (or frozing or disabling), not from chaning
bits) before source stop, so that is not like frozen. Bitmaps may be
changed in this state.
It is more like ACTIVE.

Right, it's not exactly like frozen's _implementation_ today, but...

We can move bitmap selection on the point after precopy migration, after
source stop, but I'm not sure that it would be good.

I think then we'd have four recognized states:

FROZEN: Cannot be reset/deleted. Bitmap is in-use by a block job or
other internal process. Bitmap is otherwise ACTIVE.
? Frozen means that all writes goes to the successor and frozen bitmap
itself is unchanged, no?

I was thinking from the point of view of the API. Of course internally,
you're correct; a "frozen bitmap" is one that is actually disabled and
has an anonymous successor, and that successor records IO.

  From the point of view of the API, a frozen bitmap is just "one bitmap"
that is still recording reads/writes, but is protected from being edited
from QMP.

It depends on if you're looking at bitmaps as opaque API objects or if
you're looking at the implementation.

  From an API point of view, protecting an Active bitmap from being
renamed/cleared/deleted is functionally identical to the existing case
of protecting a bitmap-and-successor pair during a backup job.
then, this should be described in API ref..

for now I see here:
# @frozen: The bitmap is currently in-use by a backup operation or block
#          and is immutable.

Which looks more like the bitmap is unchanged at all.

You're right, this is a bad wording. It's technically true of course,
but in truth the bitmap is effectively recording writes if you consider
the bitmap and the anonymous successor as "one object."

What's really immutable here from the API POV is any user-modifiable

Can we document this somehow?

In this terminology, what happens with frozen bitmap (which in fact is like "active" for the user) on successful backup finish? It turns into a difference between current-state and what-was-backed-up? It is not very transparent.. May be bitmap part, related to the backup is substituted from the bitmap?

// it's hard for me to believe that 'frozen' doesn't mean 'can not move', but if it is already established idea then ok.


Do not we want in future allow user to create successors through qmp?

I am not sure if I want to expose this functionality *DIRECTLY* but yes,
this use case will necessitate the creation of successors. I just don't
really want to give the user direct control of them via QMP. They're an
implementation detail, nothing more.

We have the following case: exteranal backup.

Backup should be done through image fleecing (temporary image, online
drive is backing for temp, start backup with sync=none from online drive
to temp, export temp through nbd).
To make this backup incremental here is BLOCK_STATUS nbd extension,
which allows dirty bitmap export.
But we need also frozen-bitmap mechanism like with normal incremental
backup.. Should we create qmp_create_bitmap_successor and friends for it?
Or we will add separate mechanism like qmp_start_external_backup,
qmp_finish_external_backup(success = true/false) ?

That was my thought at least, yes. Some kind of job mechanism that
allowed us to have start/stop/cancel mechanics that made it clear to
QEMU if it should roll back the bitmap or promote the successor.

DISABLED: Not recording any writes (by choice.)
READONLY: Not able to record any writes (by necessity.)
ACTIVE: Normal bitmap status.

Sound right?

Best regards,

reply via email to

[Prev in Thread] Current Thread [Next in Thread]