Re: [PATCH v2] fs/resctrl: Allocate rmid_ptrs[] using kvzalloc_objs()

[Reinette Chatre]

Hi Ben,

On 9/17/26 9:15 AM, Ben Horgan wrote:
> On 16/09/2026 17:52, Reinette Chatre wrote:
>> On 9/11/26 9:06 AM, Ben Horgan wrote:
>>> On an MPAM system the number of entries in rmid_ptrs[] is the number of PARTIDs
>>> multiplied by the number of PMGs. The maximum possible number of PARTIDs is
>>> 0x10000 and the maximum number of PMGs is 0x100. On a system pushing the limits
>>> of the MPAM specification this can be large enough to consistently fail the
>>> kzalloc_objs() allocation and hence fail to mount the resctrl filesystem.
>>>
>>> Switch to allocating rmid_ptrs[] using kvzalloc_objs() so that large allocations
>>> fall back to vmalloc().
>>>
>>> Fixes: 264c285999fc ("arm_mpam: resctrl: Add monitor initialisation and domain boilerplate")
>>> Signed-off-by: Ben Horgan <ben.horgan@xxxxxxx>
>>> Reviewed-by: Tony Luck <tony.luck@xxxxxxxxx>
>>> ---
>>> I found this by testing the extremes of a software model and don't expect this
>>> to effect real systems.
>>> ---
>>>  fs/resctrl/monitor.c | 4 ++--
>>>  1 file changed, 2 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/fs/resctrl/monitor.c b/fs/resctrl/monitor.c
>>> index 73413cb128ea..2d7df60973d1 100644
>>> --- a/fs/resctrl/monitor.c
>>> +++ b/fs/resctrl/monitor.c
>>> @@ -979,7 +979,7 @@ int setup_rmid_lru_list(void)
>>>  		return 0;
>>>  
>>>  	idx_limit = resctrl_arch_system_num_rmid_idx();
>>> -	rmid_ptrs = kzalloc_objs(struct rmid_entry, idx_limit);
>>> +	rmid_ptrs = kvzalloc_objs(struct rmid_entry, idx_limit);
>>>  	if (!rmid_ptrs)
>>>  		return -ENOMEM;
>>>  
>>
>> resctrl_arch_system_num_rmid_idx() that guides the rmid_ptrs[] allocation is
>> also used to size the L3 monitoring state rdt_l3_mon_domain::rmid_busy_llc[]
>> and rdt_l3_mon_domain::mbm_states[]. Each RMID is expected to have associated
>> states in these structures so to fully support this their allocation needs
>> to change also? 
> 
> At least, rdt_l3_mon_domain::mbm_states[] should use kvcalloc. I realise that I didn't hit this on
> the model I was using to test this didn't have memory bandwidth monitors.
> 
> For, rdt_l3_mon_domain::rmid_busy_llc[] the allocation will have succeeded on a model as there are
> cache occupancy counters. The allocation here is smaller than the others as it allocates a bit per
> RMID rather than multiple bytes per RMID like the other two allocations. This gives a limit here of
> 0x100 * 0x10000 / 0x8 = 0x200000 bytes, 2MB. My view here would be that, although this is quite big,
> it's not standing in the way of testing on a model and physical systems even approaching this size
> are not likely to ever be built and so doesn't justify either open coding the bitmap allocation and
> freeing or creating a new bitmap allocator. Does that make sense?

I think this may still stand in the way of testing on a model since it depends on when that testing
onlines the domains and how fragmented the memory is at that time. A simple "boot with all domains
already online" may be ok but I am not sure.

Reinette