Re: [PATCH v2 00/11] mempolicy2, mbind2, and weighted interleave

[Huang, Ying]

Gregory Price <gregory.price@xxxxxxxxxxxx> writes:

> On Tue, Dec 12, 2023 at 03:08:24PM +0800, Huang, Ying wrote:
>> Gregory Price <gregory.price@xxxxxxxxxxxx> writes:
>> 
>> >> For example, can we use something as below?
>> >> 
>> >>   long set_mempolicy2(int mode, const unsigned long *nodemask, unsigned int *il_weights,
>> >>                           unsigned long maxnode, unsigned long home_node,
>> >>                           unsigned long flags);
>> >> 
>> >>   long mbind2(unsigned long start, unsigned long len,
>> >>                           int mode, const unsigned long *nodemask, unsigned int *il_weights,
>> >>                           unsigned long maxnode, unsigned long home_node,
>> >>                           unsigned long flags);
>> >> 
>> >
>> > Your definition of mbind2 is impossible.
>> >
>> > Neither of these interfaces solve the extensibility issue.  If a new
>> > policy which requires a new format of data arrives, we can look forward
>> > to set_mempolicy3 and mbind3.
>> 
>> IIUC, we will not over-engineering too much.  It's hard to predict the
>> requirements in the future.
>> 
>
> Sure, but having the mempolicy struct at least gives us more flexibility
> than the original interface.
>
>> >> A struct may be defined to hold mempolicy iteself.
>> >> 
>> >> struct mpol {
>> >>         int mode;
>> >>         unsigned int home_node;
>> >>         const unsigned long *nodemask;
>> >>         unsigned int *il_weights;
>> >>         unsigned int maxnode;
>> >> };
>> >> 
>> >
>> > addr could be pulled out for get_mempolicy2, so i will do that
>> >
>> > 'addr_node' and 'policy_node' are warts that came from the original
>> > get_mempolicy.  Removing them increases the complexity of handling
>> > arguments in the common get_mempolicy code.
>> >
>> > I could probably just drop support for retrieving the addr_node from
>> > get_mempolicy2, since it's already possible with get_mempolicy.  So I
>> > will do that.
>> 
>> If it's necessary, we can add another struct for get_mempolicy2().  But
>> I don't think that it's necessary to add get_mempolicy2() specific
>> parameters for set_mempolicy2() or mbind2().
>
> After edits, the only parameter that doesn't have parity between
> interfaces is `addr_node` and `policy_node`.  This was an unfortunate
> wart on the original get_mempolicy() that multiplexed the output of
> (*mode) based on whether MPOL_F_NODE was set.
>
> Example:
> if (MPOL_F_ADDR | MPOL_F_NODE), then get_mempolicy() would return
> details about a VMA mempolicy + the node of that address in (*mode).
>
> Right now in get_mempolicy2() I fetch this unconditionally instead of
> requiring MPOL_F_NODE.  I did not want to multiplexing (*mode) output.
>
> I see two options:
> 1) Get rid of MPOL_F_NODE functionality in get_mempolicy2()
>    If a user wants that information, they can still use get_mempolicy()
>
> 2) Keep MPOL_F_NODE and mpol_args->addr_node/policy_node, but don't allow
>    any future extensions that create this kind of situation.

3) Add another parameter to get_mempolicy2(), such as "unsigned long
*value" to retrieve addr_node or policy_node.  We can extend it to be a
"struct *" in the future if necessary.

> I'm fine with either.  I originally aimed for get_mempolicy2() to be
> all of get_mempolicy() features + new data, but that obviously isn't
> required.

--
Best Regards,
Huang, Ying