[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [PATCH] rbd: Use RBD fast-diff for querying actual allocation
From: |
李义 |
Subject: |
Re: [PATCH] rbd: Use RBD fast-diff for querying actual allocation |
Date: |
Tue, 9 Jun 2020 21:56:35 +0800 |
On 6/9/20, Jason Dillaman <jdillama@redhat.com> wrote:
> On Tue, Jun 9, 2020 at 3:31 AM Yi Li <yili@winhong.com> wrote:
>>
>> Since Ceph version Infernalis (9.2.0) the new fast-diff mechanism
>> of RBD allows for querying actual rbd image usage.
>>
>> Prior to this version there was no easy and fast way to query how
>> much allocation a RBD image had inside a Ceph cluster.
>>
>> To use the fast-diff feature it needs to be enabled per RBD image
>> and is only supported by Ceph cluster running version Infernalis
>> (9.2.0) or newer.
>>
>> Without the fast-diff feature enabled qemu-img will report an allocation
>> identical to the image capacity.
>>
>> 'qemu-img info rbd:cepharm/liyi-rbd' might output for example:
>>
>> image: json:{"driver": "raw", "file": {"pool": "cepharm",
>> "image": "liyi-rbd", "driver": "rbd"}}
>> file format: raw
>> virtual size: 20 GiB (21474836480 bytes)
>> disk size: 0 B
>> cluster_size: 4194304
>>
>> Newly created rbds will have the fast-diff feature enabled.
>>
>> Signed-off-by: Yi Li <yili@winhong.com>
>> ---
>> block/rbd.c | 60 +++++++++++++++++++++++++++++++++++++++++++++++++++++
>> 1 file changed, 60 insertions(+)
>>
>> diff --git a/block/rbd.c b/block/rbd.c
>> index 617553b022..f231653f7b 100644
>> --- a/block/rbd.c
>> +++ b/block/rbd.c
>> @@ -1107,6 +1107,65 @@ static int64_t qemu_rbd_getlength(BlockDriverState
>> *bs)
>> return info.size;
>> }
>>
>> +#if LIBRBD_VERSION_CODE > 265
>> +static int disk_usage_callback(uint64_t offset, size_t len, int exists,
>> + void *arg)
>> +{
>> + uint64_t *used_size = (uint64_t *)(arg);
>> + if (exists) {
>> + (*used_size) += len;
>> + }
>> + return 0;
>> +}
>> +#endif
>> +
>> +static int64_t qemu_rbd_allocated_file_size(BlockDriverState *bs)
>> +{
>> + BDRVRBDState *s = bs->opaque;
>> + rbd_image_info_t info;
>> + int r;
>> + uint64_t used_size = 0;
>> + uint64_t features = 0;
>> +
>> + r = rbd_stat(s->image, &info, sizeof(info));
>> + if (r < 0) {
>> + return r;
>> + }
>> +
>> + r = rbd_get_features(s->image, &features);
>> + if (r < 0) {
>> + return r;
>> + }
>
> You should probably test the flags to ensure that the
> RBD_FLAG_FAST_DIFF_INVALID flag is not set [1]. It's potentially very
> slow and expensive to calculate the disk usage w/o a fast-diff (on
> large images) since it requires iterating over every possible 4MiB
> backing data object (by default) to query its actual usage.
>
Thanks for your review and remind me of this issue.
>> + /*
>> + * rbd_diff_iterate2() is available in versions above Ceph 0.94
>> (Hammer)
>> + * It uses a object map inside Ceph which is faster than
>> rbd_diff_iterate()
>> + * which iterates all objects.
>> + * LIBRBD_VERSION_CODE for Ceph 0.94 is 265. In 266 and upwards
>> diff_iterate2
>> + * is available
>> + */
>> +#if LIBRBD_VERSION_CODE > 265
>> + if (features & RBD_FEATURE_FAST_DIFF) {
>> +
>> + /*
>> + * RBD image fast-diff feature enabled
>> + * Querying for actual allocation.
>> + */
>> + r = rbd_diff_iterate2(s->image, NULL, 0, info.size, 0, 1,
>> + &disk_usage_callback,
>> + &used_size);
>> + if (r < 0) {
>> + return r;
>> + }
>> + } else {
>> + used_size = info.size;
>> + }
>> +#else
>> + used_size = info.size;
>> +#endif
>> + return used_size;
>> +}
>> +
>> static int coroutine_fn qemu_rbd_co_truncate(BlockDriverState *bs,
>> int64_t offset,
>> bool exact,
>> @@ -1316,6 +1375,7 @@ static BlockDriver bdrv_rbd = {
>> .bdrv_get_info = qemu_rbd_getinfo,
>> .create_opts = &qemu_rbd_create_opts,
>> .bdrv_getlength = qemu_rbd_getlength,
>> + .bdrv_get_allocated_file_size = qemu_rbd_allocated_file_size,
>> .bdrv_co_truncate = qemu_rbd_co_truncate,
>> .protocol_name = "rbd",
>>
>> --
>> 2.25.3
>>
>>
>>
>>
>
> [1]
> https://github.com/libvirt/libvirt/commit/21deeaf02fdf216b08210fc899579736973ca81d#diff-107c5451015e5980c90048ff615becb8
>
> --
> Jason
>
>