linux-kernel - Re: [PATCH v2] mm/hugetlb: defer free_huge

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Date:   Thu, 12 Dec 2019 16:01:03 -0500
From:   Waiman Long <longman@...hat.com>
To:     Andrew Morton <akpm@...ux-foundation.org>,
        Mike Kravetz <mike.kravetz@...cle.com>,
        Matthew Wilcox <willy@...radead.org>,
        linux-kernel@...r.kernel.org, linux-mm@...ck.org,
        Michal Hocko <mhocko@...nel.org>, aneesh.kumar@...ux.ibm.com
Subject: Re: [PATCH v2] mm/hugetlb: defer free_huge_page() to a workqueue

On 12/12/19 2:04 PM, Davidlohr Bueso wrote:
> There have been deadlock reports[1, 2] where put_page is called
> from softirq context and this causes trouble with the hugetlb_lock,
> as well as potentially the subpool lock.
>
> For such an unlikely scenario, lets not add irq dancing overhead
> to the lock+unlock operations, which could incur in expensive
> instruction dependencies, particularly when considering hard-irq
> safety. For example PUSHF+POPF on x86.
>
> Instead, just use a workqueue and do the free_huge_page() in regular
> task context.
>
> [1]
> https://lore.kernel.org/lkml/20191211194615.18502-1-longman@redhat.com/
> [2]
> https://lore.kernel.org/lkml/20180905112341.21355-1-aneesh.kumar@linux.ibm.com/
>
> Reported-by: Waiman Long <longman@...hat.com>
> Reported-by: Aneesh Kumar K.V <aneesh.kumar@...ux.ibm.com>
> Signed-off-by: Davidlohr Bueso <dbueso@...e.de>
> ---
>
> - Changes from v1: Only use wq when in_interrupt(), otherwise business
>    as usual. Also include the proper header file.
>
> - While I have not reproduced this issue, the v1 using wq passes all
> hugetlb
>    related tests in ltp.
>
> mm/hugetlb.c | 45 ++++++++++++++++++++++++++++++++++++++++++++-
> 1 file changed, 44 insertions(+), 1 deletion(-)
>
> diff --git a/mm/hugetlb.c b/mm/hugetlb.c
> index ac65bb5e38ac..f28cf601938d 100644
> --- a/mm/hugetlb.c
> +++ b/mm/hugetlb.c
> @@ -27,6 +27,7 @@
> #include <linux/swapops.h>
> #include <linux/jhash.h>
> #include <linux/numa.h>
> +#include <linux/workqueue.h>
>
> #include <asm/page.h>
> #include <asm/pgtable.h>
> @@ -1136,7 +1137,13 @@ static inline void
> ClearPageHugeTemporary(struct page *page)
>     page[2].mapping = NULL;
> }
>
> -void free_huge_page(struct page *page)
> +static struct workqueue_struct *hugetlb_free_page_wq;
> +struct hugetlb_free_page_work {
> +    struct page *page;
> +    struct work_struct work;
> +};
> +
> +static void __free_huge_page(struct page *page)
> {
>     /*
>      * Can't pass hstate in here because it is called from the
> @@ -1199,6 +1206,36 @@ void free_huge_page(struct page *page)
>     spin_unlock(&hugetlb_lock);
> }
>
> +static void free_huge_page_workfn(struct work_struct *work)
> +{
> +    struct page *page;
> +
> +    page = container_of(work, struct hugetlb_free_page_work,
> work)->page;
> +    __free_huge_page(page);
> +}
> +
> +void free_huge_page(struct page *page)
> +{
> +    if (unlikely(in_interrupt())) {

in_interrupt() also include context where softIRQ is disabled. So maybe
!in_task() is a better fit here.


> +        /*
> +         * While uncommon, free_huge_page() can be at least
> +         * called from softirq context, defer freeing such
> +         * that the hugetlb_lock and spool->lock need not have
> +         * to deal with irq dances just for this.
> +         */
> +        struct hugetlb_free_page_work work;
> +
> +        work.page = page;
> +        INIT_WORK_ONSTACK(&work.work, free_huge_page_workfn);
> +        queue_work(hugetlb_free_page_wq, &work.work);
> +
> +        /* wait until the huge page freeing is done */
> +        flush_work(&work.work);
> +        destroy_work_on_stack(&work.work);

The problem I see is that you don't want to wait too long while in the
hardirq context. However, the latency for the work to finish is
indeterminate.

Cheers,
Longman