percpu memory setup and num_possible_cpus() vs. nr_cpu_ids

[Jan Beulich]

Tejun,

this code of yours, as I learned the hard way during some Xen kernel work,
assumes that cpu_possible_map is not sparse. While on native x86 this is
certainly true, I wonder whether such a dependency should really exist in
architecture neutral code. Below, for reference, the places I needed to
change (4k allocator only). However, there are more instances of this in the
other allocators - those seem more problematic, as they directly drive input
to alloc_bootmem() (hence a lot of memory could be wasted, or allocation
could even fail, with a very sparse cpu_possible_map if one would simply
also use nr_cpu_ids here). In the remap (x86 - under the assumption that
if this change gets made, it should also be made in x86 code, albeit not
directly affected) allocator's case things may not be that bad, since the
memory gets freed again at the end of setup_pcpu_remap().

Thoughts?

Jan

--- head-2009-05-19.orig/arch/x86/kernel/setup_percpu.c 2009-05-19 09:55:40.000000000 +0200
+++ head-2009-05-19/arch/x86/kernel/setup_percpu.c 2009-05-27 11:12:46.000000000 +0200
@@ -295,13 +295,14 @@ static ssize_t __init setup_pcpu_4k(size
pcpu4k_nr_static_pages = PFN_UP(static_size);

/* unaligned allocations can't be freed, round up to page size */
- pages_size = PFN_ALIGN(pcpu4k_nr_static_pages * num_possible_cpus()
+ pages_size = PFN_ALIGN(pcpu4k_nr_static_pages * nr_cpu_ids
* sizeof(pcpu4k_pages[0]));
pcpu4k_pages = alloc_bootmem(pages_size);

/* allocate and copy */
j = 0;
- for_each_possible_cpu(cpu)
+ for_each_possible_cpu(cpu) {
+ j = cpu * pcpu4k_nr_static_pages;
for (i = 0; i < pcpu4k_nr_static_pages; i++) {
void *ptr;

@@ -312,6 +313,7 @@ static ssize_t __init setup_pcpu_4k(size
memcpy(ptr, __per_cpu_load + i * PAGE_SIZE, PAGE_SIZE);
pcpu4k_pages[j++] = virt_to_page(ptr);
}
+ }

/* we're ready, commit */
pr_info("PERCPU: Allocated %d 4k pages, static data %zu bytes\n",
@@ -323,8 +325,11 @@ static ssize_t __init setup_pcpu_4k(size
goto out_free_ar;

enomem:
- while (--j >= 0)
+ while (--j >= 0) {
+ if (!pcpu4k_pages[j])
+ continue;
free_bootmem(__pa(page_address(pcpu4k_pages[j])), PAGE_SIZE);
+ }
ret = -ENOMEM;
out_free_ar:
free_bootmem(__pa(pcpu4k_pages), pages_size);
--- head-2009-05-19.orig/mm/percpu.c 2009-04-21 10:18:29.000000000 +0200
+++ head-2009-05-19/mm/percpu.c 2009-05-27 11:34:01.000000000 +0200
@@ -604,7 +604,7 @@ static void pcpu_free_area(struct pcpu_c
static void pcpu_unmap(struct pcpu_chunk *chunk, int page_start, int page_end,
bool flush)
{
- unsigned int last = num_possible_cpus() - 1;
+ unsigned int last = nr_cpu_ids - 1;
unsigned int cpu;

/* unmap must not be done on immutable chunk */
@@ -690,7 +690,7 @@ static void pcpu_depopulate_chunk(struct
*/
static int pcpu_map(struct pcpu_chunk *chunk, int page_start, int page_end)
{
- unsigned int last = num_possible_cpus() - 1;
+ unsigned int last = nr_cpu_ids - 1;
unsigned int cpu;
int err;

@@ -1115,9 +1115,9 @@ size_t __init pcpu_setup_first_chunk(pcp
PFN_UP(size_sum));

pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
- pcpu_chunk_size = num_possible_cpus() * pcpu_unit_size;
+ pcpu_chunk_size = nr_cpu_ids * pcpu_unit_size;
pcpu_chunk_struct_size = sizeof(struct pcpu_chunk)
- + num_possible_cpus() * pcpu_unit_pages * sizeof(struct page *);
+ + nr_cpu_ids * pcpu_unit_pages * sizeof(struct page *);

if (dyn_size < 0)
dyn_size = pcpu_unit_size - static_size - reserved_size;

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majo...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

[Tejun Heo]

Hello, Jan.

Jan Beulich wrote:
> Tejun,
>
> this code of yours, as I learned the hard way during some Xen kernel work,
> assumes that cpu_possible_map is not sparse. While on native x86 this is
> certainly true, I wonder whether such a dependency should really exist in
> architecture neutral code. Below, for reference, the places I needed to
> change (4k allocator only). However, there are more instances of this in the
> other allocators - those seem more problematic, as they directly drive input
> to alloc_bootmem() (hence a lot of memory could be wasted, or allocation
> could even fail, with a very sparse cpu_possible_map if one would simply
> also use nr_cpu_ids here). In the remap (x86 - under the assumption that
> if this change gets made, it should also be made in x86 code, albeit not
> directly affected) allocator's case things may not be that bad, since the
> memory gets freed again at the end of setup_pcpu_remap().

There's other related larger issue. It looks like possible but
offline configuration can be quite common with virtualization and
allocating for all possible cpus on boot can waste problematic amount
of space, so the plan is to allocate only for online cpus on boot and
allocate the rest as cpus come online. This will require audit of
percpu users and probably a new cpumask representing cpus which have
ever been up.

The issue you're describing should be solved when the above is
implemented, right?

Thanks.

--
tejun