[08/12] reopen: add a transaction to drain_end nodes picked in bdrv_reopen_parse_file_or_backing

Depending on the options given to reopen_state,
bdrv_reopen_parse_file_or_backing could pick another bs
that could be from another graph, and thus not protected
by subtree_drained_begin called by the callers of this
function.

We can't simply drain-undrain here, because of transactions.
To simplify the logic, transactions always assume that they
are run under drain, so the various subtree_drain introduced
so far always take care of covering tran_commit().

And since we cannot directly do it, as the transaction is
created/committed higher above, we can just add a new
transaction to the list that just executes subtree_drained_end
to match the drained_begin done in this function.

Signed-off-by: Emanuele Giuseppe Esposito <eesposit@redhat.com>
---
 block.c | 25 ++++++++++++++++++++-----
 1 file changed, 20 insertions(+), 5 deletions(-)

On 1/18/22 17:27, Emanuele Giuseppe Esposito wrote:
> Depending on the options given to reopen_state,
> bdrv_reopen_parse_file_or_backing could pick another bs
> that could be from another graph, and thus not protected
> by subtree_drained_begin called by the callers of this
> function.
> 
> We can't simply drain-undrain here, because of transactions.
> To simplify the logic, transactions always assume that they
> are run under drain, so the various subtree_drain introduced
> so far always take care of covering tran_commit().
> 
> And since we cannot directly do it, as the transaction is
> created/committed higher above, we can just add a new
> transaction to the list that just executes subtree_drained_end
> to match the drained_begin done in this function.

A rewrite of the last two paragraphs:

---
Of the two callers of bdrv_set_file_or_backing_noperm, 
bdrv_reopen_parse_file_or_backing is unique in that it does not complete 
the reopen, it only prepares a transaction.  The actual commit (or 
abort) is done higher in the call chain.

Therefore, the call to bdrv_subtree_drained_end_unlocked must be delayed 
until after the transaction's fate has been determined and its actions 
have been performed.  Do this by recording a TransactionActionDrv to end 
the drained section for new_child_bs.
---

Likewise, the subject can be "block: keep children drained across a 
reopen transaction"; explain the change and not the technicalities of 
how you did it.

Paolo

> Signed-off-by: Emanuele Giuseppe Esposito <eesposit@redhat.com>
> ---
>   block.c | 25 ++++++++++++++++++++-----
>   1 file changed, 20 insertions(+), 5 deletions(-)
> 
> diff --git a/block.c b/block.c
> index fb5bc3077a..fcc44a49a0 100644
> --- a/block.c
> +++ b/block.c
> @@ -4522,6 +4522,10 @@ int bdrv_reopen_set_read_only(BlockDriverState *bs, bool read_only,
>       return bdrv_reopen(bs, opts, true, errp);
>   }
>   
> +TransactionActionDrv bdrv_drv_subtree_end = {
> +    .clean = (void (*)(void *)) bdrv_subtree_drained_end_unlocked,
> +};
> +
>   /*
>    * Take a BDRVReopenState and check if the value of 'backing' in the
>    * reopen_state->options QDict is valid or not.
> @@ -4550,6 +4554,7 @@ static int bdrv_reopen_parse_file_or_backing(BDRVReopenState *reopen_state,
>       const char *child_name = is_backing ? "backing" : "file";
>       QObject *value;
>       const char *str;
> +    int ret = 0;
>   
>       assert(qemu_in_main_thread());
>   
> @@ -4573,6 +4578,8 @@ static int bdrv_reopen_parse_file_or_backing(BDRVReopenState *reopen_state,
>                          "cycle", str, child_name, bs->node_name);
>               return -EINVAL;
>           }
> +        /* This will be paired with a drained_end in tran_commit */
> +        bdrv_subtree_drained_begin_unlocked(new_child_bs);
>           break;
>       default:
>           /*
> @@ -4583,18 +4590,19 @@ static int bdrv_reopen_parse_file_or_backing(BDRVReopenState *reopen_state,
>       }
>   
>       if (old_child_bs == new_child_bs) {
> -        return 0;
> +        goto end;
>       }
>   
>       if (old_child_bs) {
>           if (bdrv_skip_implicit_filters(old_child_bs) == new_child_bs) {
> -            return 0;
> +            goto end;
>           }
>   
>           if (old_child_bs->implicit) {
>               error_setg(errp, "Cannot replace implicit %s child of %s",
>                          child_name, bs->node_name);
> -            return -EPERM;
> +            ret = -EPERM;
> +            goto end;
>           }
>       }
>   
> @@ -4605,7 +4613,8 @@ static int bdrv_reopen_parse_file_or_backing(BDRVReopenState *reopen_state,
>            */
>           error_setg(errp, "'%s' is a %s filter node that does not support a "
>                      "%s child", bs->node_name, bs->drv->format_name, child_name);
> -        return -EINVAL;
> +        ret = -EINVAL;
> +        goto end;
>       }
>   
>       if (is_backing) {
> @@ -4614,8 +4623,14 @@ static int bdrv_reopen_parse_file_or_backing(BDRVReopenState *reopen_state,
>           reopen_state->old_file_bs = old_child_bs;
>       }
>   
> -    return bdrv_set_file_or_backing_noperm(bs, new_child_bs, is_backing,
> +    ret =  bdrv_set_file_or_backing_noperm(bs, new_child_bs, is_backing,
>                                              tran, errp);
> +
> +end:
> +    if (new_child_bs) {
> +        tran_add(tran, &bdrv_drv_subtree_end, new_child_bs);
> +    }
> +    return ret;
>   }
>   
>   /*

On Tue, Jan 18, 2022 at 11:27:34AM -0500, Emanuele Giuseppe Esposito wrote:
> Depending on the options given to reopen_state,
> bdrv_reopen_parse_file_or_backing could pick another bs
> that could be from another graph, and thus not protected
> by subtree_drained_begin called by the callers of this
> function.
> 
> We can't simply drain-undrain here, because of transactions.
> To simplify the logic, transactions always assume that they
> are run under drain, so the various subtree_drain introduced
> so far always take care of covering tran_commit().
> 
> And since we cannot directly do it, as the transaction is
> created/committed higher above, we can just add a new
> transaction to the list that just executes subtree_drained_end
> to match the drained_begin done in this function.
> 
> Signed-off-by: Emanuele Giuseppe Esposito <eesposit@redhat.com>
> ---
>  block.c | 25 ++++++++++++++++++++-----
>  1 file changed, 20 insertions(+), 5 deletions(-)
> 
> diff --git a/block.c b/block.c
> index fb5bc3077a..fcc44a49a0 100644
> --- a/block.c
> +++ b/block.c
> @@ -4522,6 +4522,10 @@ int bdrv_reopen_set_read_only(BlockDriverState *bs, bool read_only,
>      return bdrv_reopen(bs, opts, true, errp);
>  }
>  
> +TransactionActionDrv bdrv_drv_subtree_end = {
> +    .clean = (void (*)(void *)) bdrv_subtree_drained_end_unlocked,

Please don't cast function pointers. If the types don't match please
define a wrapper function so the compiler can check the types.