netfilter: nft_set_pipapo: speed up bulk element insertions

Insertions into the set are slow when we try to add many elements.
For 800k elements I get:

time nft -f pipapo_800k
real    19m34.849s
user    0m2.390s
sys     19m12.828s

perf stats:
 --95.39%--nft_pipapo_insert
     |--76.60%--pipapo_insert
     |           --76.37%--pipapo_resize
     |                     |--72.87%--memcpy_orig
     |                     |--1.88%--__free_pages_ok
     |                     |          --0.89%--free_tail_page_prepare
     |                      --1.38%--kvmalloc_node
     ..
     --18.56%--pipapo_get.isra.0
     |--13.91%--__bitmap_and
     |--3.01%--pipapo_refill
     |--0.81%--__kmalloc
     |           --0.74%--__kmalloc_large_node
     |                      --0.66%--__alloc_pages
     ..
     --0.52%--memset_orig

So lots of time is spent in copying exising elements to make space for
the next one.

Instead of allocating to the exact size of the new rule count, allocate
extra slack to reduce alloc/copy/free overhead.

After:
time nft -f pipapo_800k
real    1m54.110s
user    0m2.515s
sys     1m51.377s

 --80.46%--nft_pipapo_insert
     |--73.45%--pipapo_get.isra.0
     |--57.63%--__bitmap_and
     |          |--8.52%--pipapo_refill
     |--3.45%--__kmalloc
     |           --3.05%--__kmalloc_large_node
     |                      --2.58%--__alloc_pages
     --2.59%--memset_orig
     |--6.51%--pipapo_insert
            --5.96%--pipapo_resize
                     |--3.63%--memcpy_orig
                     --2.13%--kvmalloc_node

The new @rules_alloc fills a hole, so struct size doesn't go up.
Also make it so rule removal doesn't shrink unless the free/extra space
exceeds two pages.  This should be safe as well:

When a rule gets removed, the attempt to lower the allocated size is
already allowed to fail.

Exception: do exact allocations as long as set is very small (less
than one page needed).

v2: address comments from Stefano:
    kdoc comment
    formatting changes
    remove redundant assignment
    switch back to PAGE_SIZE

Link: https://lore.kernel.org/netfilter-devel/20240213141753.17ef27a6@elisabeth/
Reviewed-by: Stefano Brivio <sbrivio@redhat.com>
Signed-off-by: Florian Westphal <fw@strlen.de>

1 files changed, 2 insertions, 0 deletions

diff --git a/net/netfilter/nft_set_pipapo.h b/net/netfilter/nft_set_pipapo.h
index ca64f7505e714..24cd1ff73f981 100644
--- a/net/netfilter/nft_set_pipapo.h
+++ b/net/netfilter/nft_set_pipapo.h
@@ -106,6 +106,7 @@ union nft_pipapo_map_bucket {
  * struct nft_pipapo_field - Lookup, mapping tables and related data for a field
  * @rules:	Number of inserted rules
  * @bsize:	Size of each bucket in lookup table, in longs
+ * @rules_alloc: Number of allocated rules, always >= rules
  * @groups:	Amount of bit groups
  * @bb:		Number of bits grouped together in lookup table buckets
  * @lt:		Lookup table: 'groups' rows of buckets
@@ -114,6 +115,7 @@ union nft_pipapo_map_bucket {
 struct nft_pipapo_field {
 	unsigned int rules;
 	unsigned int bsize;
+	unsigned int rules_alloc;
 	u8 groups;
 	u8 bb;
 	unsigned long *lt;