Re: [PATCH] mm/vmalloc: fix numa spreading for large hash tables

[Nicholas Piggin]

Excerpts from Chen Wandun's message of October 14, 2021 6:59 pm:
> 
> 
> 在 2021/10/14 5:46, Shakeel Butt 写道:
>> On Tue, Sep 28, 2021 at 5:03 AM Chen Wandun <chenwandun@xxxxxxxxxx> wrote:
>>>
>>> Eric Dumazet reported a strange numa spreading info in [1], and found
>>> commit 121e6f3258fe ("mm/vmalloc: hugepage vmalloc mappings") introduced
>>> this issue [2].
>>>
>>> Dig into the difference before and after this patch, page allocation has
>>> some difference:
>>>
>>> before:
>>> alloc_large_system_hash
>>>      __vmalloc
>>>          __vmalloc_node(..., NUMA_NO_NODE, ...)
>>>              __vmalloc_node_range
>>>                  __vmalloc_area_node
>>>                      alloc_page /* because NUMA_NO_NODE, so choose alloc_page branch */
>>>                          alloc_pages_current
>>>                              alloc_page_interleave /* can be proved by print policy mode */
>>>
>>> after:
>>> alloc_large_system_hash
>>>      __vmalloc
>>>          __vmalloc_node(..., NUMA_NO_NODE, ...)
>>>              __vmalloc_node_range
>>>                  __vmalloc_area_node
>>>                      alloc_pages_node /* choose nid by nuam_mem_id() */
>>>                          __alloc_pages_node(nid, ....)
>>>
>>> So after commit 121e6f3258fe ("mm/vmalloc: hugepage vmalloc mappings"),
>>> it will allocate memory in current node instead of interleaving allocate
>>> memory.
>>>
>>> [1]
>>> https://lore.kernel.org/linux-mm/CANn89iL6AAyWhfxdHO+jaT075iOa3XcYn9k6JJc7JR2XYn6k_Q@xxxxxxxxxxxxxx/
>>>
>>> [2]
>>> https://lore.kernel.org/linux-mm/CANn89iLofTR=AK-QOZY87RdUZENCZUT4O6a0hvhu3_EwRMerOg@xxxxxxxxxxxxxx/
>>>
>>> Fixes: 121e6f3258fe ("mm/vmalloc: hugepage vmalloc mappings")
>>> Reported-by: Eric Dumazet <edumazet@xxxxxxxxxx>
>>> Signed-off-by: Chen Wandun <chenwandun@xxxxxxxxxx>
>>> ---
>>>   mm/vmalloc.c | 33 ++++++++++++++++++++++++++-------
>>>   1 file changed, 26 insertions(+), 7 deletions(-)
>>>
>>> diff --git a/mm/vmalloc.c b/mm/vmalloc.c
>>> index f884706c5280..48e717626e94 100644
>>> --- a/mm/vmalloc.c
>>> +++ b/mm/vmalloc.c
>>> @@ -2823,6 +2823,8 @@ vm_area_alloc_pages(gfp_t gfp, int nid,
>>>                  unsigned int order, unsigned int nr_pages, struct page **pages)
>>>   {
>>>          unsigned int nr_allocated = 0;
>>> +       struct page *page;
>>> +       int i;
>>>
>>>          /*
>>>           * For order-0 pages we make use of bulk allocator, if
>>> @@ -2833,6 +2835,7 @@ vm_area_alloc_pages(gfp_t gfp, int nid,
>>>          if (!order) {
>> 
>> Can you please replace the above with if (!order && nid != NUMA_NO_NODE)?
>> 
>>>                  while (nr_allocated < nr_pages) {
>>>                          unsigned int nr, nr_pages_request;
>>> +                       page = NULL;
>>>
>>>                          /*
>>>                           * A maximum allowed request is hard-coded and is 100
>>> @@ -2842,9 +2845,23 @@ vm_area_alloc_pages(gfp_t gfp, int nid,
>>>                           */
>>>                          nr_pages_request = min(100U, nr_pages - nr_allocated);
>>>
>> 
>> Undo the following change in this if block.
> 
> Yes, It seem like more simpler as you suggested, But it still have 
> performance regression, I plan to change the following to consider
> both mempolcy and alloc_pages_bulk.

Thanks for finding and debugging this. These APIs are a maze of twisty 
little passages, all alike so I could be as confused as I was when I
wrote that patch, but doesn't a minimal fix look something like this?

diff --git a/mm/vmalloc.c b/mm/vmalloc.c
index d77830ff604c..75ee9679f521 100644
--- a/mm/vmalloc.c
+++ b/mm/vmalloc.c
@@ -2860,7 +2860,10 @@ vm_area_alloc_pages(gfp_t gfp, int nid,
                struct page *page;
                int i;
 
-               page = alloc_pages_node(nid, gfp, order);
+               if (nid == NUMA_NO_NODE)
+                       page = alloc_pages(gfp, order);
+               else
+                       page = alloc_pages_node(nid, gfp, order);
                if (unlikely(!page))
                        break;
 

Thanks,
Nick