Re: [PATCH RFC net-next 3/3] mm: make zone->free_area[order] access faster
From: Mel Gorman
Date: Thu Feb 25 2021 - 06:30:47 EST
As a side-node, I didn't pick up the other patches as there is review
feedback and I didn't have strong opinions either way. Patch 3 is curious
though, it probably should be split out and sent separetly but still;
On Wed, Feb 24, 2021 at 07:56:51PM +0100, Jesper Dangaard Brouer wrote:
>
Avoid multiplication (imul) operations when accessing:
>
zone->free_area[order].nr_free
>
>
This was really tricky to find. I was puzzled why perf reported that
>
rmqueue_bulk was using 44% of the time in an imul operation:
>
>
??? del_page_from_free_list():
>
44,54 ??? e2: imul 0ドルx58,%rax,%rax
>
>
This operation was generated (by compiler) because the struct free_area have
>
size 88 bytes or 0x58 hex. The compiler cannot find a shift operation to use
>
and instead choose to use a more expensive imul, to find the offset into the
>
array free_area[].
>
>
The patch align struct free_area to a cache-line, which cause the
>
compiler avoid the imul operation. The imul operation is very fast on
>
modern Intel CPUs. To help fast-path that decrement 'nr_free' move the
>
member 'nr_free' to be first element, which saves one 'add' operation.
>
>
Looking up instruction latency this exchange a 3-cycle imul with a
>
1-cycle shl, saving 2-cycles. It does trade some space to do this.
>
>
Used: gcc (GCC) 9.3.1 20200408 (Red Hat 9.3.1-2)
>
I'm having some trouble parsing this and matching it to the patch itself.
First off, on my system (x86-64), the size of struct free area is 72,
not 88 bytes. For either size, cache-aligning the structure is a big
increase in the struct size.
struct free_area {
struct list_head free_list[4]; /* 0 64 */
/* --- cacheline 1 boundary (64 bytes) --- */
long unsigned int nr_free; /* 64 8 */
/* size: 72, cachelines: 2, members: 2 */
/* last cacheline: 8 bytes */
};
Are there other patches in the tree? What does pahole say?
With gcc-9, I'm also not seeing the imul instruction outputted like you
described in rmqueue_pcplist which inlines rmqueue_bulk. At the point
where it calls get_page_from_free_area, it's using shl for the page list
operation. This might be a compiler glitch but given that free_area is a
different size, I'm less certain and wonder if something else is going on.
Finally, moving nr_free to the end and cache aligning it will make the
started of each free_list cache-aligned because of its location in the
struct zone so what purpose does __pad_to_align_free_list serve?
--
Mel Gorman
SUSE Labs