Re: [PATCH v7 10/14] xen: add cache coloring allocator for domains

[Jan Beulich <jbeulich@suse.com>]

On 15.03.2024 11:58, Carlo Nonato wrote:
> Add a new memory page allocator that implements the cache coloring mechanism.
> The allocation algorithm enforces equal frequency distribution of cache
> partitions, following the coloring configuration of a domain. This allows
> for an even utilization of cache sets for every domain.
> 
> Pages are stored in a color-indexed array of lists. Those lists are filled
> by a simple init function which computes the color of each page.
> When a domain requests a page, the allocator extract the page from the list
> with the maximum number of free pages between those that the domain can
> access, given its coloring configuration.

Minor remark: I'm not a native speaker, but "between" here reads odd to
me. I'd have expected perhaps "among".

> --- a/docs/misc/xen-command-line.pandoc
> +++ b/docs/misc/xen-command-line.pandoc
> @@ -270,6 +270,20 @@ and not running softirqs. Reduce this if softirqs are not being run frequently
>  enough. Setting this to a high value may cause boot failure, particularly if
>  the NMI watchdog is also enabled.
>  
> +### buddy-alloc-size (arm64)
> +> `= <size>`
> +
> +> Default: `64M`
> +
> +Amount of memory reserved for the buddy allocator when colored allocator is
> +active. This options is parsed only when LLC coloring support is enabled.

Nit: s/parsed/used/ - the option is always parsed as long as LLC_COLORING=y.

> @@ -1945,6 +1949,164 @@ static unsigned long avail_heap_pages(
>      return free_pages;
>  }
>  
> +/*************************
> + * COLORED SIDE-ALLOCATOR
> + *
> + * Pages are grouped by LLC color in lists which are globally referred to as the
> + * color heap. Lists are populated in end_boot_allocator().
> + * After initialization there will be N lists where N is the number of
> + * available colors on the platform.
> + */
> +static struct page_list_head *__ro_after_init _color_heap;
> +#define color_heap(color) (&_color_heap[color])
> +
> +static unsigned long *__ro_after_init free_colored_pages;
> +
> +/* Memory required for buddy allocator to work with colored one */
> +#ifdef CONFIG_LLC_COLORING
> +static unsigned long __initdata buddy_alloc_size =
> +    MB(CONFIG_BUDDY_ALLOCATOR_SIZE);
> +size_param("buddy-alloc-size", buddy_alloc_size);
> +
> +#define domain_num_llc_colors(d) (d)->num_llc_colors
> +#define domain_llc_color(d, i)   (d)->llc_colors[i]
> +#else
> +static unsigned long __initdata buddy_alloc_size;
> +
> +#define domain_num_llc_colors(d) 0
> +#define domain_llc_color(d, i)   0
> +#endif
> +
> +static void free_color_heap_page(struct page_info *pg, bool need_scrub)
> +{
> +    unsigned int color = page_to_llc_color(pg);
> +    struct page_list_head *head = color_heap(color);
> +
> +    spin_lock(&heap_lock);
> +
> +    mark_page_free(pg, page_to_mfn(pg));
> +
> +    if ( need_scrub )
> +    {
> +        pg->count_info |= PGC_need_scrub;
> +        poison_one_page(pg);
> +    }
> +
> +    free_colored_pages[color]++;
> +    page_list_add(pg, head);

May I please ask for a comment (or at least some wording in the description)
as to the choice made here between head or tail insertion? When assuming
that across a system there's no sharing of colors, preferably re-using
cache-hot pages is certainly good. Whereas when colors can reasonably be
expected to be shared, avoiding to quickly re-use a freed page can also
have benefits.

> +static struct page_info *alloc_color_heap_page(unsigned int memflags,
> +                                               const struct domain *d)
> +{
> +    struct page_info *pg = NULL;
> +    unsigned int i, color = 0;
> +    unsigned long max = 0;
> +    bool need_tlbflush = false;
> +    uint32_t tlbflush_timestamp = 0;
> +    bool need_scrub;
> +
> +    if ( memflags >> _MEMF_bits )
> +        return NULL;

By mentioning MEMF_bits earlier on I meant to give an example. What
about MEMF_node and in particular MEMF_exact_node? Certain other flags
also aren't obvious as to being okay to silently ignore.

> +    spin_lock(&heap_lock);
> +
> +    for ( i = 0; i < domain_num_llc_colors(d); i++ )
> +    {
> +        unsigned long free = free_colored_pages[domain_llc_color(d, i)];
> +
> +        if ( free > max )
> +        {
> +            color = domain_llc_color(d, i);
> +            pg = page_list_first(color_heap(color));
> +            max = free;
> +        }
> +    }
> +
> +    if ( !pg )
> +    {
> +        spin_unlock(&heap_lock);
> +        return NULL;
> +    }
> +
> +    need_scrub = pg->count_info & (PGC_need_scrub);
> +    pg->count_info = PGC_state_inuse | (pg->count_info & PGC_colored);

Better PGC_preserved?

> +static void __init init_color_heap_pages(struct page_info *pg,
> +                                         unsigned long nr_pages)
> +{
> +    unsigned int i;
> +    bool need_scrub = opt_bootscrub == BOOTSCRUB_IDLE;
> +
> +    if ( buddy_alloc_size )
> +    {
> +        unsigned long buddy_pages = min(PFN_DOWN(buddy_alloc_size), nr_pages);
> +
> +        init_heap_pages(pg, buddy_pages);

There's a corner case where init_heap_pages() would break when passed 0
as 2nd argument. I think you want to alter the enclosing if() to
"if ( buddy_alloc_size >= PAGE_SIZE )" to be entirely certain to avoid
that case.

> +static void dump_color_heap(void)
> +{
> +    unsigned int color;
> +
> +    printk("Dumping color heap info\n");
> +    for ( color = 0; color < get_max_nr_llc_colors(); color++ )
> +        if ( free_colored_pages[color] > 0 )
> +            printk("Color heap[%u]: %lu pages\n",
> +                   color, free_colored_pages[color]);
> +}

While having all of the code above from here outside of any #ifdef is
helpful to prevent unintended breakage when changes are made and tested
only on non-Arm64 targets, I'd still like to ask: Halfway recent
compilers manage to eliminate everything? I'd like to avoid e.g. x86
being left with traces of coloring despite not being able at all to use
it.

> @@ -2485,7 +2660,10 @@ struct page_info *alloc_domheap_pages(
>          }
>          if ( assign_page(pg, order, d, memflags) )
>          {
> -            free_heap_pages(pg, order, memflags & MEMF_no_scrub);
> +            if ( pg->count_info & PGC_colored )
> +                free_color_heap_page(pg, memflags & MEMF_no_scrub);
> +            else
> +                free_heap_pages(pg, order, memflags & MEMF_no_scrub);
>              return NULL;
>          }
>      }
> @@ -2568,7 +2746,10 @@ void free_domheap_pages(struct page_info *pg, unsigned int order)
>              scrub = 1;
>          }
>  
> -        free_heap_pages(pg, order, scrub);
> +        if ( pg->count_info & PGC_colored )
> +            free_color_heap_page(pg, scrub);
> +        else
> +            free_heap_pages(pg, order, scrub);
>      }

Instead of this, did you consider altering free_heap_pages() to forward
to free_color_heap_page()? That would then also allow to have a single,
central comment and/or assertion that the "order" value here isn't lost.

Jan