Fix race in concurrent_vector::grow_by()
#1532
Conversation
Signed-off-by: Fedotov, Aleksei <[email protected]>
Signed-off-by: Fedotov, Aleksei <[email protected]>
Signed-off-by: Fedotov, Aleksei <[email protected]>
There was a problem hiding this comment.
Other than this, the patch looks good for me.
I would still wait for feedback from @pavelkumbrasev to ensure we did not miss something.
| } else if (new_table) { | ||
| // Other thread was the first to replace the segment table. Current thread's | ||
| // table is not needed anymore, so destroying it. | ||
| destroy_and_deallocate_table(new_table, pointers_per_long_table); | ||
| } | ||
| }).on_exception([&] { | ||
| my_segment_table_allocation_failed.store(true, std::memory_order_relaxed); |
There was a problem hiding this comment.
Would this logic work correct while having one thread which allocates the long table and successfully stores it into the my_segment_table and another which allocates the long table and receives a bad_alloc while trying to allocate?
It seems like we need to double check this here.
There was a problem hiding this comment.
It depends on what logic we think is correct. Perhaps, not reporting allocation error in case it succeeded by some other thread makes sense, and double checking my_segment_table for updated state first can improve algorithm robustness in some cases. Although, I am not sure how frequent these cases can be. Also, it is not full cover anyway, as successful allocation can still sneak in after that double check. Full cover requires recording the failure state into single source of information such as my_segment_table or some kind of non-trivial synchronization between multiple information sources, but we don't have that kind of logic through the whole vector right now I guess. So, at most we can write something like the following instead:
| my_segment_table_allocation_failed.store(true, std::memory_order_relaxed); | |
| // Last chance to overcome the failure, hoping that other thread has succeeded | |
| // extending the table | |
| table = get_table(); | |
| if (table == my_embedded_table) { | |
| my_segment_table_allocation_failed.store(true, std::memory_order_relaxed); | |
| } |
Does it suffice though?
Description
When thread tries to grow
concurrent_vector, it first creates locally a new extended segmentation table, with which it then tries to update the global one. However, before this patch, the thread simply updated the global table not taking into account that there might be the other threads doing the same thing. This patch makes the thread aware about its possibly concurrent environment.Fixes #1531
Type of change
Choose one or multiple, leave empty if none of the other choices apply
Add a respective label(s) to PR if you have permissions
Tests
Documentation
Breaks backward compatibility
Notify the following users
@npotravkin
Other information