[lkml]   [2018]   [May]   [22]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [PATCH v3] mlx4_core: allocate ICM memory in page size chunks

On 5/22/2018 8:33 AM, Tariq Toukan wrote:
> On 18/05/2018 12:45 AM, Qing Huang wrote:
>> On 5/17/2018 2:14 PM, Eric Dumazet wrote:
>>> On 05/17/2018 01:53 PM, Qing Huang wrote:
>>>> When a system is under memory presure (high usage with fragments),
>>>> the original 256KB ICM chunk allocations will likely trigger kernel
>>>> memory management to enter slow path doing memory compact/migration
>>>> ops in order to complete high order memory allocations.
>>>> When that happens, user processes calling uverb APIs may get stuck
>>>> for more than 120s easily even though there are a lot of free pages
>>>> in smaller chunks available in the system.
>>>> Syslog:
>>>> ...
>>>> Dec 10 09:04:51 slcc03db02 kernel: [397078.572732] INFO: task
>>>> oracle_205573_e:205573 blocked for more than 120 seconds.
>>>> ...
>>> NACK on this patch.
>>> You have been asked repeatedly to use kvmalloc()
>>> This is not a minor suggestion.
>>> Take a look
>>> at
>> Would you please take a look at how table->icm is being used in the
>> mlx4 driver? It's a meta data used for individual pointer variable
>> referencing,
>> not as data frag or in/out buffer. It has no need for contiguous phy.
>> memory.
>> Thanks.
> This would cause a degradation when iterating the entries of table->icm.
> For example, in mlx4_table_get_range.
int mlx4_table_get_range(struct mlx4_dev *dev, struct mlx4_icm_table *table,
                         u32 start, u32 end)
        int inc = MLX4_TABLE_CHUNK_SIZE / table->obj_size;
        int err;
        u32 i;

        for (i = start; i <= end; i += inc) {
                err = mlx4_table_get(dev, table, i);
                if (err)
                        goto fail;

        return 0;

E.g. mtt obj is 8 bytes, so a 4KB ICM block would have 512 mtt objects.
So you will have to allocate
more 512 mtt objects in order to have table->icm pointer to increment by
1 to fetch next pointer
value.  So 256K mtt objects are needed in order to traverse table->icm
pointer across a page boundary
in the call stacks.

Considering mlx4_table_get_range() is only used in control path, there
is no significant gain by using kvzalloc
vs. vzalloc for table->icm.

Anyway, if a user makes sure mlx4 driver to be loaded very early and
doesn't remove and reload it afterwards,
we should have enough (and not wasting) contiguous phy mem for
table->icm allocation. I will use kvzalloc to
replace vzalloc and send a V4 patch.


> Thanks,
> Tariq
>>> And you'll understand some people care about this.
>>> Strongly.
>>> Thanks.
> --
> To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
> the body of a message to
> More majordomo info at

 \ /
  Last update: 2018-05-23 03:43    [W:0.059 / U:2.184 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site