Re: [PATCH v8 3/5] ipc: Allow boot time extension of IPCMNI from 32k to 2M

[Waiman Long]

On 10/02/2018 12:32 PM, Manfred Spraul wrote:
> Hello together,
>
> On 8/18/18 3:15 AM, Waiman Long wrote:
>> On 08/17/2018 12:45 PM, Davidlohr Bueso wrote:
>>> Cc'ing Manfred.
>>>
>>> On Mon, 18 Jun 2018, Waiman Long wrote:
>>>
>>>> The maximum number of unique System V IPC identifiers was limited to
>>>> 32k.  That limit should be big enough for most use cases.
>>>>
>>>> However, there are some users out there requesting for more. To
>>>> satisfy
>>>> the need of those users, a new boot time kernel option "ipcmni_extend"
>>>> is added to extend the IPCMNI value to 2M. This is a 64X increase
>>>> which
>>>> hopefully is big enough for them.
>>> Could you please provide more info on the need of these users and how
>>> you came up with this new value (which just seems quite arbitrary)?
>>>
>>> Thanks,
>>> Davidlohr
>> Red Hat has a customer that is migrating from Solaris to Linux. Some of
>> their applications just happen to use more than 32k of shared memory
>> segments. I think Solaris allows up to 16M unique ID.
>>
>> Yes, the amount of increase is a bit arbitrary. I was trying to balance
>> how many bits should be left for sequence number. Maybe I should just
>> take 8 more bits for ID and leave 8 bits for sequence number to match
>> Solaris.
>
> - I think we should use the same numbers as Solaris.
> Otherwise we later have to touch it again.

As said in my patch, it is a trade-off between # of uniq identifiers
versus the chance of id reuse. I don't have an objection to increase it
further, but I don't see the customers to really need such a large value.

>
> - What is the performance when using shmget() with already 10M
> segments present?

I am not sure the performance impact as I had not measure it myself. The
shmget() function is considered in slowpath. We are generally less
concern about its performance than other code paths that are in a
performance critical path.

>
> - I like the new logic for updating the sequence counter.
>
> Is there a reason why you only enable it for extended mode?

I tried not to disturb the existing logic for backward compatibility
concern. I don't mind switching it all over to use the new "deleted"
approach if other people have no objection.

Cheers,
Longman

> You create a rarely used codepath, and I don't understand what speaks
> against switching to the 'deleted' approach for all systems.
>
>
> -- 
>
>     Manfred
>