[Tarantool-patches] [PATCH v3 1/2] b-tree: return NULL on matras_alloc fail

[Konstantin Osipov]

* Nikita Pettik <korablev at tarantool.org> [20/02/04 21:10]:
> On 04 Feb 20:25, Konstantin Osipov wrote:
> > * Nikita Pettik <korablev at tarantool.org> [20/02/04 20:21]:
> > > > > On 20 Jan 21:13, Ilya Kosarev wrote:
> > > > > > In bps_tree_create_leaf we use matras_alloc in case
> > > > > > bps_tree_garbage_pop didn't work out. However it also might not
> > > > > > succeed. Then we need to return NULL instead of dereferencing NULL
> > > > > > pointer.
> > > > 
> > > > I don't understand the attempt to fix it.
> > > > 
> > > > The reason the allocations are not checked - most likely -because
> > > > BPS should refuse to even begin an operation if there is not
> > > > enough memory in matras.
> > > 
> > > According to the code it doesn't look so. Matras allocation checked
> > > on fails, except two ones in bps_tree_create_inner() and
> > > bps_tree_create_leaf().
> > 
> > Because reserve_blocks() should reserve enough blocks or fail, and
> > it is called before create_inner/create_leaf in all execution paths?
> 
> Consider following path:
> 
> memtx_tree_index_replace
> |
> ->memtx_tree_insert (bps_tree_insert)
>   |
>   ->bps_tree_insert_first_elem
>     |
>     ->bps_tree_create_leaf
>       |
>       -> matras_alloc
> 
> In this case reserve_blocks() is not called. AFAIU path is likely to
> be reachable. Am I missing smth?

No, in this trace I agree with you, we clearly see some sort of
refactoring artefact or a coding bug.

What I am observing is that bps_tree_insert_first_elem() actually checks
the return value of bps_tree_create_leaf(). But
bps_tree_create_leaf() never returns NULL, so this check is never
false.

bps was originally written in C++, so all of these checks for OOm
were added when it was rewritten in C, after the fact.

Looks like we're dealing with an artefact of this rewrite - the
checks are not consistent.


-- 
Konstantin Osipov, Moscow, Russia
https://scylladb.com