Re: [PATCH V2] block/rbd: implement bdrv_co_block_status

[Ilya Dryomov <idryomov@gmail.com>]

On Thu, Sep 2, 2021 at 4:12 PM Peter Lieven <pl@kamp.de> wrote:
>
> Am 24.08.21 um 22:39 schrieb Ilya Dryomov:
> > On Mon, Aug 23, 2021 at 11:38 AM Peter Lieven <pl@kamp.de> wrote:
> >> Am 22.08.21 um 23:02 schrieb Ilya Dryomov:
> >>> On Tue, Aug 10, 2021 at 3:41 PM Peter Lieven <pl@kamp.de> wrote:
> >>>> the qemu rbd driver currently lacks support for bdrv_co_block_status.
> >>>> This results mainly in incorrect progress during block operations (e.g.
> >>>> qemu-img convert with an rbd image as source).
> >>>>
> >>>> This patch utilizes the rbd_diff_iterate2 call from librbd to detect
> >>>> allocated and unallocated (all zero areas).
> >>>>
> >>>> To avoid querying the ceph OSDs for the answer this is only done if
> >>>> the image has the fast-diff features which depends on the object-map
> >>> Hi Peter,
> >>>
> >>> Nit: "has the fast-diff feature which depends on the object-map and
> >>> exclusive-lock features"
> >>
> >> will reword in V3.
> >>
> >>
> >>>> and exclusive-lock. In this case it is guaranteed that the information
> >>>> is present in memory in the librbd client and thus very fast.
> >>>>
> >>>> If fast-diff is not available all areas are reported to be allocated
> >>>> which is the current behaviour if bdrv_co_block_status is not implemented.
> >>>>
> >>>> Signed-off-by: Peter Lieven <pl@kamp.de>
> >>>> ---
> >>>> V1->V2:
> >>>> - add commit comment [Stefano]
> >>>> - use failed_post_open [Stefano]
> >>>> - remove redundant assert [Stefano]
> >>>> - add macro+comment for the magic -9000 value [Stefano]
> >>>> - always set *file if its non NULL [Stefano]
> >>>>
> >>>>    block/rbd.c | 125 ++++++++++++++++++++++++++++++++++++++++++++++++++++
> >>>>    1 file changed, 125 insertions(+)
> >>>>
> >>>> diff --git a/block/rbd.c b/block/rbd.c
> >>>> index dcf82b15b8..8692e76f40 100644
> >>>> --- a/block/rbd.c
> >>>> +++ b/block/rbd.c
> >>>> @@ -88,6 +88,7 @@ typedef struct BDRVRBDState {
> >>>>        char *namespace;
> >>>>        uint64_t image_size;
> >>>>        uint64_t object_size;
> >>>> +    uint64_t features;
> >>>>    } BDRVRBDState;
> >>>>
> >>>>    typedef struct RBDTask {
> >>>> @@ -983,6 +984,13 @@ static int qemu_rbd_open(BlockDriverState *bs, QDict *options, int flags,
> >>>>        s->image_size = info.size;
> >>>>        s->object_size = info.obj_size;
> >>>>
> >>>> +    r = rbd_get_features(s->image, &s->features);
> >>>> +    if (r < 0) {
> >>>> +        error_setg_errno(errp, -r, "error getting image features from %s",
> >>>> +                         s->image_name);
> >>>> +        goto failed_post_open;
> >>>> +    }
> >>> The object-map and fast-diff features can be enabled/disabled while the
> >>> image is open so this should probably go to qemu_rbd_co_block_status().
> >>>
> >>>> +
> >>>>        /* If we are using an rbd snapshot, we must be r/o, otherwise
> >>>>         * leave as-is */
> >>>>        if (s->snap != NULL) {
> >>>> @@ -1259,6 +1267,122 @@ static ImageInfoSpecific *qemu_rbd_get_specific_info(BlockDriverState *bs,
> >>>>        return spec_info;
> >>>>    }
> >>>>
> >>>> +typedef struct rbd_diff_req {
> >>>> +    uint64_t offs;
> >>>> +    uint64_t bytes;
> >>>> +    int exists;
> >>>> +} rbd_diff_req;
> >>>> +
> >>>> +/*
> >>>> + * rbd_diff_iterate2 allows to interrupt the exection by returning a negative
> >>>> + * value in the callback routine. Choose a value that does not conflict with
> >>>> + * an existing exitcode and return it if we want to prematurely stop the
> >>>> + * execution because we detected a change in the allocation status.
> >>>> + */
> >>>> +#define QEMU_RBD_EXIT_DIFF_ITERATE2 -9000
> >>>> +
> >>>> +static int qemu_rbd_co_block_status_cb(uint64_t offs, size_t len,
> >>>> +                                       int exists, void *opaque)
> >>>> +{
> >>>> +    struct rbd_diff_req *req = opaque;
> >>>> +
> >>>> +    assert(req->offs + req->bytes <= offs);
> >>>> +
> >>>> +    if (req->exists && offs > req->offs + req->bytes) {
> >>>> +        /*
> >>>> +         * we started in an allocated area and jumped over an unallocated area,
> >>>> +         * req->bytes contains the length of the allocated area before the
> >>>> +         * unallocated area. stop further processing.
> >>>> +         */
> >>>> +        return QEMU_RBD_EXIT_DIFF_ITERATE2;
> >>>> +    }
> >>>> +    if (req->exists && !exists) {
> >>>> +        /*
> >>>> +         * we started in an allocated area and reached a hole. req->bytes
> >>>> +         * contains the length of the allocated area before the hole.
> >>>> +         * stop further processing.
> >>>> +         */
> >>>> +        return QEMU_RBD_EXIT_DIFF_ITERATE2;
> >>>> +    }
> >>>> +    if (!req->exists && exists && offs > req->offs) {
> >>>> +        /*
> >>>> +         * we started in an unallocated area and hit the first allocated
> >>>> +         * block. req->bytes must be set to the length of the unallocated area
> >>>> +         * before the allocated area. stop further processing.
> >>>> +         */
> >>>> +        req->bytes = offs - req->offs;
> >>>> +        return QEMU_RBD_EXIT_DIFF_ITERATE2;
> >>>> +    }
> >>>> +
> >>>> +    /*
> >>>> +     * assert that we catched all cases above and allocation state has not
> >>> catched -> caught
> >>>
> >>>> +     * changed during callbacks.
> >>>> +     */
> >>>> +    assert(exists == req->exists || !req->bytes);
> >>>> +    req->exists = exists;
> >>>> +
> >>>> +    /*
> >>>> +     * assert that we either return an unallocated block or have got callbacks
> >>>> +     * for all allocated blocks present.
> >>>> +     */
> >>>> +    assert(!req->exists || offs == req->offs + req->bytes);
> >>>> +    req->bytes = offs + len - req->offs;
> >>>> +
> >>>> +    return 0;
> >>>> +}
> >>>> +
> >>>> +static int coroutine_fn qemu_rbd_co_block_status(BlockDriverState *bs,
> >>>> +                                                 bool want_zero, int64_t offset,
> >>>> +                                                 int64_t bytes, int64_t *pnum,
> >>>> +                                                 int64_t *map,
> >>>> +                                                 BlockDriverState **file)
> >>>> +{
> >>>> +    BDRVRBDState *s = bs->opaque;
> >>>> +    int ret, r;
> >>>> +    struct rbd_diff_req req = { .offs = offset };
> >>>> +
> >>>> +    assert(offset + bytes <= s->image_size);
> >>>> +
> >>>> +    /* default to all sectors allocated */
> >>>> +    ret = BDRV_BLOCK_DATA | BDRV_BLOCK_OFFSET_VALID;
> >>> I'm a little confused by the meaning of these flags (but I haven't
> >>> looked at the other drivers yet).  Looks like this patch always sets
> >>> BDRV_BLOCK_OFFSET_VALID (makes sense since the "host" offset is always
> >>> known for rbd) and returns either BDRV_BLOCK_DATA or BDRV_BLOCK_ZERO.
> >>>
> >>> DATA ZERO OFFSET_VALID
> >>>    t    t        t       sectors read as zero, returned file is zero at offset
> >>>    t    f        t       sectors read as valid from file at offset
> >>>    f    t        t       sectors preallocated, read as zero, returned file not
> >>>                          necessarily zero at offset
> >>>    f    f        t       sectors preallocated but read from backing_hd,
> >>>                          returned file contains garbage at offset
> >>>
> >>> What about the first case (BDRV_BLOCK_DATA | BDRV_BLOCK_ZERO)?
> >>> What is the practical difference to just BDRV_BLOCK_ZERO?
> >>
> >> Actually I don't know, I adapted the flags from other drivers. Looking at your
> >> table it seems that DATA + ZERO would be more appropriate for rbd, right?
> > It would seem that way but I haven't had a chance to look at how qemu
> > block layer interprets these yet.
> >
> >> We do not preallocate in qem/rbd yet.
> > I guess it depends on what preallocated means here, if it talks about
> > the logical address space or the physical sectors.  If the former, rbd
> > can be thought of as always preallocated.
> >
> > Separately, rbd_read() at an offset that corresponds to an unallocated
> > area (including in parent image(s)) would always return zeroes.  I wonder
> > if that is the same as "returned file is zero at offset".
>
>
> It seems that all other drivers either set DATA or ZERO. I would leave it
>
> the same for RBD.

Agreed.  It looks like it has to do with layering and backing files and
since rbd layering isn't exposed to qemu let's not over complicate this.

>
>
> >
> >>
> >>>> +    if (map) {
> >>>> +        *map = offset;
> >>>> +    }
> >>>> +    if (file) {
> >>>> +        *file = bs;
> >>>> +    }
> >>> A comment in block_int.h says that map and file are guaranteed to be
> >>> non-NULL so these tests seem redundant?
> >>
> >> This is also Copy&Paste from other drivers, will change it (and maybe
> >>
> >> in other drivers as well).
> >>
> >>
> >>
> >>>> +    *pnum = bytes;
> >>>> +
> >>>> +    /* RBD image does not support fast-diff */
> >>>> +    if (!(s->features & RBD_FEATURE_FAST_DIFF)) {
> >>>> +        goto out;
> >>>> +    }
> >>> Need to make sure that fast-diff is valid here: call rbd_get_flags()
> >>> on the image and test for !RBD_FLAG_FAST_DIFF_INVALID.
> >>
> >> Do I have to check for feature FAST_DIFF + flag !FAST_DIFF_INVALID or
> >>
> >> is the second enough? Is this call fast?
> > For both.  The call should be as fast as rbd_get_features().
>
>
> The question was if this is handled solely in the client?

Yes, most of the time.  Both image features and flags are stored in
the header object and the in-memory copy of the header may need to be
refreshed if the client gets a header update notification in the
interim.

Thanks,

                Ilya