lkml.org 
[lkml]   [2018]   [Feb]   [15]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [RFC PATCH V2 00/22] Intel(R) Resource Director Technology Cache Pseudo-Locking enabling
From
Date
On 02/15/2018 12:39 PM, Reinette Chatre wrote:
> On 2/14/2018 10:31 AM, Reinette Chatre wrote:
>> On 2/14/2018 10:12 AM, Mike Kravetz wrote:
>>> On 02/13/2018 07:46 AM, Reinette Chatre wrote:
>>>> Adding MM maintainers to v2 to share the new MM change (patch 21/22) that
>>>> enables large contiguous regions that was created to support large Cache
>>>> Pseudo-Locked regions (patch 22/22). This week MM team received another
>>>> proposal to support large contiguous allocations ("[RFC PATCH 0/3]
>>>> Interface for higher order contiguous allocations" at
>>>> http://lkml.kernel.org/r/20180212222056.9735-1-mike.kravetz@oracle.com).
>>>> I have not yet tested with this new proposal but it does seem appropriate
>>>> and I should be able to rework patch 22 from this series on top of that if
>>>> it is accepted instead of what I have in patch 21 of this series.
>>>>
>>>
>>> Well, I certainly would prefer the adoption and use of a more general
>>> purpose interface rather than exposing alloc_gigantic_page().
>>>
>>> Both the interface I suggested and alloc_gigantic_page end up calling
>>> alloc_contig_range(). I have not looked at your entire patch series, but
>>> do be aware that in its present form alloc_contig_range will run into
>>> issues if called by two threads simultaneously for the same page range.
>>> Calling alloc_gigantic_page without some form of synchronization will
>>> expose this issue. Currently this is handled by hugetlb_lock for all
>>> users of alloc_gigantic_page. If you simply expose alloc_gigantic_page
>>> without any type of synchronization, you may run into issues. The first
>>> patch in my RFC "mm: make start_isolate_page_range() fail if already
>>> isolated" should handle this situation IF we decide to expose
>>> alloc_gigantic_page (which I do not suggest).
>>
>> My work depends on the ability to create large contiguous regions,
>> creating these large regions is not the goal in itself. Certainly I
>> would want to use the most appropriate mechanism and I would gladly
>> modify my work to do so.
>>
>> I do not insist on using alloc_gigantic_page(). Now that I am aware of
>> your RFC I started the process to convert to the new
>> find_alloc_contig_pages(). I did not do so earlier because it was not
>> available when I prepared this work for submission. I plan to respond to
>> your RFC when my testing is complete but please give me a few days to do
>> so. Could you please also cc me if you do send out any new versions?
>
> Testing with the new find_alloc_contig_pages() introduced in
> "[RFC PATCH 0/3] Interface for higher order contiguous allocations" at
> http://lkml.kernel.org/r/20180212222056.9735-1-mike.kravetz@oracle.com
> was successful. If this new interface is merged then Cache
> Pseudo-Locking can easily be ported to use that instead of what I have
> in patch 21/22 (exposing alloc_gigantic_page()) with the following
> change to patch 22/22:
>

Nice. Thank you for converting and testing with this interface.
--
Mike Kravetz

>
> diff --git a/arch/x86/kernel/cpu/intel_rdt_pseudo_lock.c
> b/arch/x86/kernel/cpu/intel_rdt_pseudo_lock.c
> index 99918943a98a..b5e4ae379352 100644
> --- a/arch/x86/kernel/cpu/intel_rdt_pseudo_lock.c
> +++ b/arch/x86/kernel/cpu/intel_rdt_pseudo_lock.c
> @@ -228,9 +228,10 @@ static int contig_mem_alloc(struct
> pseudo_lock_region *plr)
> }
>
> if (plr->size > KMALLOC_MAX_SIZE) {
> - plr->kmem = alloc_gigantic_page(cpu_to_node(plr->cpu),
> - get_order(plr->size),
> - GFP_KERNEL | __GFP_ZERO);
> + plr->kmem = find_alloc_contig_pages(get_order(plr->size),
> + GFP_KERNEL | __GFP_ZERO,
> + cpu_to_node(plr->cpu),
> + NULL);
> if (!plr->kmem) {
> rdt_last_cmd_puts("unable to allocate gigantic
> page\n");
> return -ENOMEM;
> @@ -255,7 +256,7 @@ static int contig_mem_alloc(struct
> pseudo_lock_region *plr)
> static void contig_mem_free(struct pseudo_lock_region *plr)
> {
> if (plr->size > KMALLOC_MAX_SIZE)
> - free_gigantic_page(plr->kmem, get_order(plr->size));
> + free_contig_pages(plr->kmem, 1 << get_order(plr->size));
> else
> kfree(page_to_virt(plr->kmem));
> }
>
>
> It does seem as though there will be a new API for large contiguous
> allocations, eliminating the need for patch 21 of this series. How large
> contiguous regions are allocated are independent of Cache Pseudo-Locking
> though and the patch series as submitted still stands. I can include the
> above snippet in a new version of the series but I am not sure if it is
> preferred at this time. Please do let me know, I'd be happy to.
>
> Reinette
>
> --
> To unsubscribe, send a message with 'unsubscribe linux-mm' in
> the body to majordomo@kvack.org. For more info on Linux MM,
> see: http://www.linux-mm.org/ .
> Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>
>

\
 
 \ /
  Last update: 2018-02-15 22:11    [W:0.072 / U:8.424 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site