Re: [PATCH v2] resource: Merge resources on a node when hot-adding memory

[Rashmica Gupta <rashmica.g@gmail.com>]

On Tue, Aug 7, 2018 at 9:52 PM, Vlastimil Babka <vbabka@suse.cz> wrote:
> On 08/06/2018 08:52 AM, Rashmica Gupta wrote:
>> When hot-removing memory release_mem_region_adjustable() splits
>> iomem resources if they are not the exact size of the memory being
>> hot-deleted. Adding this memory back to the kernel adds a new
>> resource.
>>
>> Eg a node has memory 0x0 - 0xfffffffff. Offlining and hot-removing
>> 1GB from 0xf40000000 results in the single resource 0x0-0xfffffffff being
>> split into two resources: 0x0-0xf3fffffff and 0xf80000000-0xfffffffff.
>>
>> When we hot-add the memory back we now have three resources:
>> 0x0-0xf3fffffff, 0xf40000000-0xf7fffffff, and 0xf80000000-0xfffffffff.
>>
>> Now if we try to remove a section of memory that overlaps these resources,
>> like 2GB from 0xf40000000, release_mem_region_adjustable() fails as it
>> expects the chunk of memory to be within the boundaries of a single
>> resource.
>
> Hi,
>
> it's the first time I see the resource code, so I might be easily wrong.
> How can it happen that the second remove is section aligned but the
> first one not?
>

I probably shouldn't have used that word... When I said "a section of memory"
I really meant "a chunk of memory" or "some memory".


>> This patch adds a function request_resource_and_merge(). This is called
>> instead of request_resource_conflict() when registering a resource in
>> add_memory(). It calls request_resource_conflict() and if hot-removing is
>> enabled (if it isn't we won't get resource fragmentation) we attempt to
>> merge contiguous resources on the node.
>>
>> Signed-off-by: Rashmica Gupta <rashmica.g@gmail.com>
> ...
>> --- a/kernel/resource.c
>> +++ b/kernel/resource.c
> ...
>> +/*
>> + * Attempt to merge resources on the node
>> + */
>> +static void merge_node_resources(int nid, struct resource *parent)
>> +{
>> +     struct resource *res;
>> +     uint64_t start_addr;
>> +     uint64_t end_addr;
>> +     int ret;
>> +
>> +     start_addr = node_start_pfn(nid) << PAGE_SHIFT;
>> +     end_addr = node_end_pfn(nid) << PAGE_SHIFT;
>> +
>> +     write_lock(&resource_lock);
>> +
>> +     /* Get the first resource */
>> +     res = parent->child;
>> +
>> +     while (res) {
>> +             /* Check that the resource is within the node */
>> +             if (res->start < start_addr) {
>> +                     res = res->sibling;
>> +                     continue;
>> +             }
>> +             /* Exit if resource is past end of node */
>> +             if (res->sibling->end > end_addr)
>> +                     break;
>
> IIUC, resource end is closed, so adjacent resources's start is end+1.
> But node_end_pfn is open, so the comparison above should use '>='
> instead of '>'?

You are right. Thanks for spotting that.

>
>> +
>> +             ret = merge_resources(res);
>> +             if (!ret)
>> +                     continue;
>> +             res = res->sibling;
>
> Should this rather use next_resource() to merge at all levels of the
> hierarchy? Although memory seems to be flat under &iomem_resource so it
> would be just future-proofing.

I don't know enough about the hierarchy and layout of resources to comment on
this.

>
> Thanks,
> Vlastimil