Yet another optimization is to extend the mirroring iteration to include more
adjacent dirty blocks. This limits the number of I/O operations and makes
mirroring efficient even with a small granularity. Most of the infrastructure
is already in place; we only need to put a loop around the computation of
the origin and sector count of the iteration.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
With AIO support in place, we can start copying more than one chunk
in parallel. This patch introduces the required infrastructure for
this: the buffer is split into multiple granularity-sized chunks,
and there is a free list to access them.
Because of copy-on-write, a single operation may already require
multiple chunks to be available on the free list.
In addition, two different iterations on the HBitmap may want to
copy the same cluster. We avoid this by keeping a bitmap of in-flight
I/O operations, and blocking until the previous iteration completes.
This should be a pretty rare occurrence, though; as long as there is
no overlap the next iteration can start before the previous one finishes.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
This makes sense when the next commit starts using the extra buffer space
to perform many I/O operations asynchronously.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
There is really no change in the behavior of the job here, since
there is still a maximum of one in-flight I/O operation between
the source and the target. However, this patch already introduces
the AIO callbacks (which are unmodified in the next patch)
and some of the logic to count in-flight operations and only
complete the job when there is none.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
The desired granularity may be very different depending on the kind of
operation (e.g. continuous replication vs. collapse-to-raw) and whether
the VM is expected to perform lots of I/O while mirroring is in progress.
Allow the user to customize it, while providing a sane default so that
in general there will be no extra allocated space in the target compared
to the source.
Reviewed-by: Eric Blake <eblake@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Laszlo Ersek <lersek@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
When mirroring runs, the backing files for the target may not yet be
ready. However, this means that a copy-on-write operation on the target
would fill the missing sectors with zeros. Copy-on-write only happens
if the granularity of the dirty bitmap is smaller than the cluster size
(and only for clusters that are allocated in the source after the job
has started copying). So far, the granularity was fixed to 1MB; to avoid
the problem we detected the situation and required the backing files to
be available in that case only.
However, we want to lower the granularity for efficiency, so we need
a better solution. The solution is to always copy a whole cluster the
first time it is touched. The code keeps a bitmap of clusters that
have already been allocated by the mirroring job, and only does "manual"
copy-on-write if the chunk being copied is zero in the bitmap.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
This actually uses the dirty bitmap in the block layer, and converts
mirroring to use an HBitmapIter.
Reviewed-by: Laszlo Ersek <lersek@redhat.com> (except block/mirror.c parts)
Reviewed-by: Eric Blake <eblake@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
It allocates with qemu_blockalign(), therefore it must free with
qemu_vfree(), not g_free().
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Error management is important for mirroring; otherwise, an error on the
target (even something as "innocent" as ENOSPC) requires to start again
with a full copy. Similar to on_read_error/on_write_error, two separate
knobs are provided for on_source_error (reads) and on_target_error (writes).
The default is 'report' for both.
The 'ignore' policy will leave the sector dirty, so that it will be
retried later. Thus, it will not cause corruption.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Switching to the target of the migration is done mostly asynchronously,
and reported to management via the BLOCK_JOB_COMPLETED event; the only
synchronous phase is opening the backing files. bdrv_open_backing_file
can always be done, even for migration of the full image (aka sync:
'full'). In this case, qmp_drive_mirror will create the target disk
with no backing file at all, and bdrv_open_backing_file will be a no-op.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
This patch adds the implementation of a new job that mirrors a disk to
a new image while letting the guest continue using the old image.
The target is treated as a "black box" and data is copied from the
source to the target in the background. This can be used for several
purposes, including storage migration, continuous replication, and
observation of the guest I/O in an external program. It is also a
first step in replacing the inefficient block migration code that is
part of QEMU.
The job is possibly never-ending, but it is logically structured into
two phases: 1) copy all data as fast as possible until the target
first gets in sync with the source; 2) keep target in sync and
ensure that reopening to the target gets a correct (full) copy
of the source data.
The second phase is indicated by the progress in "info block-jobs"
reporting the current offset to be equal to the length of the file.
When the job is cancelled in the second phase, QEMU will run the
job until the source is clean and quiescent, then it will report
successful completion of the job.
In other words, the BLOCK_JOB_CANCELLED event means that the target
may _not_ be consistent with a past state of the source; the
BLOCK_JOB_COMPLETED event means that the target is consistent with
a past state of the source. (Note that it could already happen
that management lost the race against QEMU and got a completion
event instead of cancellation).
It is not yet possible to complete the job and switch over to the target
disk. The next patches will fix this and add many refinements to the
basic idea introduced here. These include improved error management,
some tunable knobs and performance optimizations.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>