David Hildenbrand <david@xxxxxxxxxx> writes:
> On 03.04.24 13:40, Ryan Roberts wrote:
>> Multi-size THP enables performance improvements by allocating large,
>> pte-mapped folios for anonymous memory. However I've observed that on an
>> arm64 system running a parallel workload (e.g. kernel compilation)
>> across many cores, under high memory pressure, the speed regresses. This
>> is due to bottlenecking on the increased number of TLBIs added due to
>> all the extra folio splitting when the large folios are swapped out.
>> Therefore, solve this regression by adding support for swapping out
>> mTHP
>> without needing to split the folio, just like is already done for
>> PMD-sized THP. This change only applies when CONFIG_THP_SWAP is enabled,
>> and when the swap backing store is a non-rotating block device. These
>> are the same constraints as for the existing PMD-sized THP swap-out
>> support.
>> Note that no attempt is made to swap-in (m)THP here - this is still
>> done
>> page-by-page, like for PMD-sized THP. But swapping-out mTHP is a
>> prerequisite for swapping-in mTHP.
>> The main change here is to improve the swap entry allocator so that
>> it
>> can allocate any power-of-2 number of contiguous entries between [1, (1
>> << PMD_ORDER)]. This is done by allocating a cluster for each distinct
>> order and allocating sequentially from it until the cluster is full.
>> This ensures that we don't need to search the map and we get no
>> fragmentation due to alignment padding for different orders in the
>> cluster. If there is no current cluster for a given order, we attempt to
>> allocate a free cluster from the list. If there are no free clusters, we
>> fail the allocation and the caller can fall back to splitting the folio
>> and allocates individual entries (as per existing PMD-sized THP
>> fallback).
>> The per-order current clusters are maintained per-cpu using the
>> existing
>> infrastructure. This is done to avoid interleving pages from different
>> tasks, which would prevent IO being batched. This is already done for
>> the order-0 allocations so we follow the same pattern.
>> As is done for order-0 per-cpu clusters, the scanner now can steal
>> order-0 entries from any per-cpu-per-order reserved cluster. This
>> ensures that when the swap file is getting full, space doesn't get tied
>> up in the per-cpu reserves.
>> This change only modifies swap to be able to accept any order
>> mTHP. It
>> doesn't change the callers to elide doing the actual split. That will be
>> done in separate changes.
>> Reviewed-by: "Huang, Ying" <ying.huang@xxxxxxxxx>
>> Signed-off-by: Ryan Roberts <ryan.roberts@xxxxxxx>
>> ---
>> include/linux/swap.h | 10 ++-
>> mm/swap_slots.c | 6 +-
>> mm/swapfile.c | 175 ++++++++++++++++++++++++-------------------
>> 3 files changed, 109 insertions(+), 82 deletions(-)
>> diff --git a/include/linux/swap.h b/include/linux/swap.h
>> index 5e1e4f5bf0cb..11c53692f65f 100644
>> --- a/include/linux/swap.h
>> +++ b/include/linux/swap.h
>> @@ -268,13 +268,19 @@ struct swap_cluster_info {
>> */
>> #define SWAP_NEXT_INVALID 0
>> +#ifdef CONFIG_THP_SWAP
>> +#define SWAP_NR_ORDERS (PMD_ORDER + 1)
>> +#else
>> +#define SWAP_NR_ORDERS 1
>> +#endif
>> +
>> /*
>> * We assign a cluster to each CPU, so each CPU can allocate swap entry from
>> * its own cluster and swapout sequentially. The purpose is to optimize swapout
>> * throughput.
>> */
>> struct percpu_cluster {
>> - unsigned int next; /* Likely next allocation offset */
>> + unsigned int next[SWAP_NR_ORDERS]; /* Likely next allocation offset */
>> };
>> struct swap_cluster_list {
>> @@ -471,7 +477,7 @@ swp_entry_t folio_alloc_swap(struct folio *folio);
>> bool folio_free_swap(struct folio *folio);
>> void put_swap_folio(struct folio *folio, swp_entry_t entry);
>> extern swp_entry_t get_swap_page_of_type(int);
>> -extern int get_swap_pages(int n, swp_entry_t swp_entries[], int entry_size);
>> +extern int get_swap_pages(int n, swp_entry_t swp_entries[], int order);
>> extern int add_swap_count_continuation(swp_entry_t, gfp_t);
>> extern void swap_shmem_alloc(swp_entry_t);
>> extern int swap_duplicate(swp_entry_t);
>> diff --git a/mm/swap_slots.c b/mm/swap_slots.c
>> index 53abeaf1371d..13ab3b771409 100644
>> --- a/mm/swap_slots.c
>> +++ b/mm/swap_slots.c
>> @@ -264,7 +264,7 @@ static int refill_swap_slots_cache(struct swap_slots_cache *cache)
>> cache->cur = 0;
>> if (swap_slot_cache_active)
>> cache->nr = get_swap_pages(SWAP_SLOTS_CACHE_SIZE,
>> - cache->slots, 1);
>> + cache->slots, 0);
>> return cache->nr;
>> }
>> @@ -311,7 +311,7 @@ swp_entry_t folio_alloc_swap(struct folio *folio)
>> if (folio_test_large(folio)) {
>> if (IS_ENABLED(CONFIG_THP_SWAP))
>> - get_swap_pages(1, &entry, folio_nr_pages(folio));
>> + get_swap_pages(1, &entry, folio_order(folio));
>
> The only comment I have is that this nr_pages -> order conversion adds
> a bit of noise to this patch.
>
> AFAIKS, it's primarily only required for "cluster->next[order]",
> everything else doesn't really require the order.
>
> I'd just have split that out into a separate patch, or simply
> converted nr_pages -> order where required.
>
> Nothing jumped at me, but I'm not an expert on that code, so I'm
> mostly trusting the others ;)
The nr_pages -> order conversion replaces ilog2(nr_pages) with
(1<<order). IIUC, "<<" is a little faster than "ilog2()". And, we
don't need to worry about whether nr_pages is a power of 2. Do you
think that this makes sense?
--
Best Regards,
Huang, Ying
