On 2024/7/16 20:58, Yunsheng Lin wrote:
...
>
> Option 1 assuming nc->remaining as a negative value does not seems to
> make it a more maintainable solution than option 2. How about something
> like below if using a negative value to enable some optimization like LEA
> does not have a noticeable performance difference?
Suppose the below as option 3, it seems the option 3 has better performance
than option 2, and option 2 has better performance than option 1 using the
ko introduced in patch 1.
Option 1:
Performance counter stats for 'insmod ./page_frag_test.ko test_push_cpu=16 test_pop_cpu=17 test_alloc_len=12 nr_test=5120000' (500 runs):
17.757768 task-clock (msec) # 0.001 CPUs utilized ( +- 0.17% )
5 context-switches # 0.288 K/sec ( +- 0.28% )
0 cpu-migrations # 0.007 K/sec ( +- 12.36% )
82 page-faults # 0.005 M/sec ( +- 0.06% )
46128280 cycles # 2.598 GHz ( +- 0.17% )
60938595 instructions # 1.32 insn per cycle ( +- 0.02% )
14783794 branches # 832.525 M/sec ( +- 0.02% )
20393 branch-misses # 0.14% of all branches ( +- 0.13% )
24.556644680 seconds time elapsed ( +- 0.07% )
Option 2:
Performance counter stats for 'insmod ./page_frag_test.ko test_push_cpu=16 test_pop_cpu=17 test_alloc_len=12 nr_test=5120000' (500 runs):
18.443508 task-clock (msec) # 0.001 CPUs utilized ( +- 0.61% )
6 context-switches # 0.342 K/sec ( +- 0.57% )
0 cpu-migrations # 0.025 K/sec ( +- 4.89% )
82 page-faults # 0.004 M/sec ( +- 0.06% )
47901207 cycles # 2.597 GHz ( +- 0.61% )
60985019 instructions # 1.27 insn per cycle ( +- 0.05% )
14787177 branches # 801.755 M/sec ( +- 0.05% )
21099 branch-misses # 0.14% of all branches ( +- 0.14% )
24.413183804 seconds time elapsed ( +- 0.06% )
Option 3:
Performance counter stats for 'insmod ./page_frag_test.ko test_push_cpu=16 test_pop_cpu=17 test_alloc_len=12 nr_test=5120000' (500 runs):
17.847031 task-clock (msec) # 0.001 CPUs utilized ( +- 0.23% )
5 context-switches # 0.305 K/sec ( +- 0.55% )
0 cpu-migrations # 0.017 K/sec ( +- 6.86% )
82 page-faults # 0.005 M/sec ( +- 0.06% )
46355974 cycles # 2.597 GHz ( +- 0.23% )
60848779 instructions # 1.31 insn per cycle ( +- 0.03% )
14758941 branches # 826.969 M/sec ( +- 0.03% )
20728 branch-misses # 0.14% of all branches ( +- 0.15% )
24.376161069 seconds time elapsed ( +- 0.06% )
>
> struct page_frag_cache {
> /* encoded_va consists of the virtual address, pfmemalloc bit and order
> * of a page.
> */
> unsigned long encoded_va;
>
> #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE) && (BITS_PER_LONG <= 32)
> __u16 remaining;
> __u16 pagecnt_bias;
> #else
> __u32 remaining;
> __u32 pagecnt_bias;
> #endif
> };
>
> void *__page_frag_alloc_va_align(struct page_frag_cache *nc,
> unsigned int fragsz, gfp_t gfp_mask,
> unsigned int align_mask)
> {
> unsigned int size = page_frag_cache_page_size(nc->encoded_va);
> unsigned int remaining;
>
> remaining = nc->remaining & align_mask;
> if (unlikely(remaining < fragsz)) {
> if (unlikely(fragsz > PAGE_SIZE)) {
> /*
> * The caller is trying to allocate a fragment
> * with fragsz > PAGE_SIZE but the cache isn't big
> * enough to satisfy the request, this may
> * happen in low memory conditions.
> * We don't release the cache page because
> * it could make memory pressure worse
> * so we simply return NULL here.
> */
> return NULL;
> }
>
> if (!__page_frag_cache_refill(nc, gfp_mask))
> return NULL;
>
> size = page_frag_cache_page_size(nc->encoded_va);
> remaining = size;
> }
>
> nc->pagecnt_bias--;
> nc->remaining = remaining - fragsz;
>
> return encoded_page_address(nc->encoded_va) + (size - remaining);
> }
>
>
