Re: [Qemu-devel] [PATCH v10 2/3] quorum: implement bdrv_add_child() and bdrv_del_child()

[Changlong Xie <xiecl.fnst@cn.fujitsu.com>]

On 03/06/2016 02:13 AM, Max Reitz wrote:
> On 16.02.2016 10:37, Changlong Xie wrote:
>> From: Wen Congyang <wency@cn.fujitsu.com>
>>
>> Signed-off-by: Wen Congyang <wency@cn.fujitsu.com>
>> Signed-off-by: zhanghailiang <zhang.zhanghailiang@huawei.com>
>> Signed-off-by: Gonglei <arei.gonglei@huawei.com>
>> Signed-off-by: Changlong Xie <xiecl.fnst@cn.fujitsu.com>
>> ---
>>   block.c               |   8 ++--
>>   block/quorum.c        | 122 +++++++++++++++++++++++++++++++++++++++++++++++++-
>>   include/block/block.h |   4 ++
>>   3 files changed, 128 insertions(+), 6 deletions(-)
>>
>> diff --git a/block.c b/block.c
>> index 08aa979..c3c9dc0 100644
>> --- a/block.c
>> +++ b/block.c
>> @@ -1198,10 +1198,10 @@ static int bdrv_fill_options(QDict **options, const char *filename,
>>       return 0;
>>   }
>>
>> -static BdrvChild *bdrv_attach_child(BlockDriverState *parent_bs,
>> -                                    BlockDriverState *child_bs,
>> -                                    const char *child_name,
>> -                                    const BdrvChildRole *child_role)
>> +BdrvChild *bdrv_attach_child(BlockDriverState *parent_bs,
>> +                             BlockDriverState *child_bs,
>> +                             const char *child_name,
>> +                             const BdrvChildRole *child_role)
>>   {
>>       BdrvChild *child = g_new(BdrvChild, 1);
>>       *child = (BdrvChild) {
>> diff --git a/block/quorum.c b/block/quorum.c
>> index a5ae4b8..e5a7e4f 100644
>> --- a/block/quorum.c
>> +++ b/block/quorum.c
>> @@ -24,6 +24,7 @@
>>   #include "qapi/qmp/qstring.h"
>>   #include "qapi-event.h"
>>   #include "crypto/hash.h"
>> +#include "qemu/bitmap.h"
>>
>>   #define HASH_LENGTH 32
>>
>> @@ -81,6 +82,8 @@ typedef struct BDRVQuorumState {
>>       bool rewrite_corrupted;/* true if the driver must rewrite-on-read corrupted
>>                               * block if Quorum is reached.
>>                               */
>> +    unsigned long *index_bitmap;
>> +    int bsize;
>>
>>       QuorumReadPattern read_pattern;
>>   } BDRVQuorumState;
>> @@ -876,9 +879,9 @@ static int quorum_open(BlockDriverState *bs, QDict *options, int flags,
>>           ret = -EINVAL;
>>           goto exit;
>>       }
>> -    if (s->num_children < 2) {
>> +    if (s->num_children < 1) {
>>           error_setg(&local_err,
>> -                   "Number of provided children must be greater than 1");
>> +                   "Number of provided children must be 1 or more");
>
> Side note: Actually, we could work with 0 children, too. Quorum would
> then need to implement bdrv_is_inserted() and return false if there are
> no children.
>
> But that is something that can be implemented later on if the need arises.

Hi Max

Thanks for pointing it out.

>
>>           ret = -EINVAL;
>>           goto exit;
>>       }
>> @@ -927,6 +930,7 @@ static int quorum_open(BlockDriverState *bs, QDict *options, int flags,
>>       /* allocate the children array */
>>       s->children = g_new0(BdrvChild *, s->num_children);
>>       opened = g_new0(bool, s->num_children);
>> +    s->index_bitmap = bitmap_new(s->num_children);
>>
>>       for (i = 0; i < s->num_children; i++) {
>>           char indexstr[32];
>> @@ -942,6 +946,8 @@ static int quorum_open(BlockDriverState *bs, QDict *options, int flags,
>>
>>           opened[i] = true;
>>       }
>> +    bitmap_set(s->index_bitmap, 0, s->num_children);
>> +    s->bsize = s->num_children;
>>
>>       g_free(opened);
>>       goto exit;
>> @@ -998,6 +1004,115 @@ static void quorum_attach_aio_context(BlockDriverState *bs,
>>       }
>>   }
>>
>> +static int get_new_child_index(BDRVQuorumState *s)
>> +{
>> +    int index;
>> +
>> +    index = find_next_zero_bit(s->index_bitmap, s->bsize, 0);
>> +    if (index < s->bsize) {
>> +        return index;
>> +    }
>> +
>> +    if ((s->bsize % BITS_PER_LONG) == 0) {
>> +        s->index_bitmap = bitmap_zero_extend(s->index_bitmap, s->bsize,
>> +                                             s->bsize + 1);
>
> I think this function needs to be called unconditionally. Looking into
> its implementation, its call to g_realloc() will not do anything (and it
> will probably be pretty quick at that), but the following bitmap_clear()

Yes. If "BITS_TO_LONGS(new_nbits) == BITS_TO_LONGS(old_nbits)", 
g_realloc will do nothing.

> will only clear the bits from old_nbits (s->bsize) to new_nbits
> (s->bsize + 1).
>
> Thus, if you only call this function every 32nd/64th child, only that
> child's bit will be initialized to zero. All the rest is undefined.
>
> You probably didn't notice because bitmap_new() returns a
> zero-initialized bitmap, and thus you'd have to create around 64
> children (on an x64 machine) to notice

OOH! you're catching a *BIG* fish here. I'll remove the wrong "if" 
condition next version. *Thanks*

>
>> +    }
>> +
>> +    return s->bsize++;
>> +}
>> +
>> +static void remove_child_index(BDRVQuorumState *s, int index)
>> +{
>> +    int last_index;
>> +    long new_len;
>
> size_t would be the more appropriate type.

okay

>
>> +
>> +    assert(index < s->bsize);
>> +
>> +    clear_bit(index, s->index_bitmap);
>> +    if (index < s->bsize - 1) {
>> +        /*
>> +         * The last bit is always set, and we don't clear
>
> s/don't/didn't/

I'm going to remove "and we don't clear the last bit" here.

>
>> +         * the last bit.
>> +         */
>> +        return;
>> +    }
>> +
>> +    last_index = find_last_bit(s->index_bitmap, s->bsize);
>
> An assert(last_index < s->bsize); here wouldn't hurt.
>

okay.

> (last_index == s->bsize would be the case if no bit is set in
> s->index_bitmap anymore, which should be impossible.)
>
>> +    s->bsize = last_index + 1;
>> +    if (BITS_TO_LONGS(last_index + 1) == BITS_TO_LONGS(s->bsize)) {

I correct myself here, it should be "BITS_TO_LONGS(old_bsize) == 
BITS_TO_LONGS(s->bsize)".

>> +        return;
>> +    }
>> +
>> +    new_len = BITS_TO_LONGS(last_index + 1) * sizeof(unsigned long);
>
> s/last_index + 1/s->bsize/ looks better to me.

okay.

>
>> +    s->index_bitmap = g_realloc(s->index_bitmap, new_len);
>> +}
>> +
>> +static void quorum_add_child(BlockDriverState *bs, BlockDriverState *child_bs,
>> +                             Error **errp)
>> +{
>> +    BDRVQuorumState *s = bs->opaque;
>> +    BdrvChild *child;
>> +    char indexstr[32];
>> +    int index, ret;
>> +
>> +    index = get_new_child_index(s);
>> +    ret = snprintf(indexstr, 32, "children.%d", index);
>> +    if (ret < 0 || ret >= 32) {
>> +        error_setg(errp, "cannot generate child name");
>> +        return;
>> +    }
>> +
>> +    bdrv_drain(bs);
>> +
>> +    assert(s->num_children <= INT_MAX / sizeof(BdrvChild *));
>> +    if (s->num_children == INT_MAX / sizeof(BdrvChild *)) {
>> +        error_setg(errp, "Too many children");
>> +        return;
>> +    }
>> +    s->children = g_renew(BdrvChild *, s->children, s->num_children + 1);
>> +
>> +    bdrv_ref(child_bs);
>> +    child = bdrv_attach_child(bs, child_bs, indexstr, &child_format);
>> +    s->children[s->num_children++] = child;
>> +    set_bit(index, s->index_bitmap);
>> +}
>> +
>> +static void quorum_del_child(BlockDriverState *bs, BlockDriverState *child_bs,
>> +                             Error **errp)
>> +{
>> +    BDRVQuorumState *s = bs->opaque;
>> +    BdrvChild *child;
>> +    int i, index;
>> +
>> +    for (i = 0; i < s->num_children; i++) {
>> +        if (s->children[i]->bs == child_bs) {
>> +            break;
>> +        }
>> +    }
>> +
>> +    /* we have checked it in bdrv_del_child() */
>> +    assert(i < s->num_children);
>> +    child = s->children[i];
>> +
>> +    if (s->num_children <= s->threshold) {
>> +        error_setg(errp,
>> +            "The number of children cannot be lower than the vote threshold %d",
>> +            s->threshold);
>> +        return;
>> +    }
>> +
>> +    /* child->name is "children.%d" */
>
> Optional: assert(!strncmp(child->name, "children.", 9));
>
>> +    index = atoi(child->name + 9);
>
> Optional: Assert absence of an error:
>
> unsigned long index;
> char *endptr;
>
> index = strtoul(child->name + 9, &endptr, 10);
> assert(index >= 0 && !*endptr);

Really useful, but since we strictly named 'child->name' in 
quorum_add_child, let's just keep the orignal one.

Thanks
	-Xie
>
> Max
>
>> +
>> +    bdrv_drain(bs);
>> +    /* We can safely remove this child now */
>> +    memmove(&s->children[i], &s->children[i + 1],
>> +            (s->num_children - i - 1) * sizeof(void *));
>> +    s->children = g_renew(BdrvChild *, s->children, --s->num_children);
>> +    remove_child_index(s, index);
>> +    bdrv_unref_child(bs, child);
>> +}
>> +
>>   static void quorum_refresh_filename(BlockDriverState *bs, QDict *options)
>>   {
>>       BDRVQuorumState *s = bs->opaque;
>> @@ -1053,6 +1168,9 @@ static BlockDriver bdrv_quorum = {
>>       .bdrv_detach_aio_context            = quorum_detach_aio_context,
>>       .bdrv_attach_aio_context            = quorum_attach_aio_context,
>>
>> +    .bdrv_add_child                     = quorum_add_child,
>> +    .bdrv_del_child                     = quorum_del_child,
>> +
>>       .is_filter                          = true,
>>       .bdrv_recurse_is_first_non_filter   = quorum_recurse_is_first_non_filter,
>>   };
>> diff --git a/include/block/block.h b/include/block/block.h
>> index ecde190..4b787d2 100644
>> --- a/include/block/block.h
>> +++ b/include/block/block.h
>> @@ -517,6 +517,10 @@ void bdrv_disable_copy_on_read(BlockDriverState *bs);
>>   void bdrv_ref(BlockDriverState *bs);
>>   void bdrv_unref(BlockDriverState *bs);
>>   void bdrv_unref_child(BlockDriverState *parent, BdrvChild *child);
>> +BdrvChild *bdrv_attach_child(BlockDriverState *parent_bs,
>> +                             BlockDriverState *child_bs,
>> +                             const char *child_name,
>> +                             const BdrvChildRole *child_role);
>>
>>   bool bdrv_op_is_blocked(BlockDriverState *bs, BlockOpType op, Error **errp);
>>   void bdrv_op_block(BlockDriverState *bs, BlockOpType op, Error *reason);
>>
>
>