diff mbox series

mm/madvise: break reclaim when lock race

Message ID 20211207032202.6022-1-xiaofeng5@xiaomi.com (mailing list archive)
State New
Headers show
Series mm/madvise: break reclaim when lock race | expand

Commit Message

xf2017140389@gmail.com Dec. 7, 2021, 3:22 a.m. UTC 
From: xiaofeng <xiaofeng5@xiaomi.com>

When the process_madvise() system call is being used, it takes mmap_lock and blocks the application from allocating memory, leading to unreasonable delays. This patchset aims to fix it.

Signed-off-by: xiaofeng <xiaofeng5@xiaomi.com>
---
 mm/madvise.c | 3 +++
 1 file changed, 3 insertions(+)

Comments

Suren Baghdasaryan Dec. 7, 2021, 5:04 p.m. UTC | #1 
On Mon, Dec 6, 2021 at 7:22 PM <xf2017140389@gmail.com> wrote:
>
> From: xiaofeng <xiaofeng5@xiaomi.com>
>
> When the process_madvise() system call is being used, it takes mmap_lock and blocks the application from allocating memory, leading to unreasonable delays. This patchset aims to fix it.

Please use checkpatch.pl script before posting (see this blog for
example: http://nickdesaulniers.github.io/blog/2017/05/16/submitting-your-first-patch-to-the-linux-kernel-and-responding-to-feedback/).

Could you please describe your usecase a bit more? Which MADV is being
used specifically (MADV_COLD or MADV_PAGEOUT)?
CC'ing Minchan as he might be interested in learning more about this.

>
> Signed-off-by: xiaofeng <xiaofeng5@xiaomi.com>
> ---
>  mm/madvise.c | 3 +++
>  1 file changed, 3 insertions(+)
>
> diff --git a/mm/madvise.c b/mm/madvise.c
> index 8c927202bbe6..8f1e325873e0 100644
> --- a/mm/madvise.c
> +++ b/mm/madvise.c
> @@ -322,6 +322,9 @@ static int madvise_cold_or_pageout_pte_range(pmd_t *pmd,
>         if (fatal_signal_pending(current))
>                 return -EINTR;
>
> +       if (mmap_lock_is_contended(mm))
> +               return -EINTR;
> +
>  #ifdef CONFIG_TRANSPARENT_HUGEPAGE
>         if (pmd_trans_huge(*pmd)) {
>                 pmd_t orig_pmd;
> --
> 2.17.1
>
Matthew Wilcox Dec. 7, 2021, 5:20 p.m. UTC | #2 
On Tue, Dec 07, 2021 at 11:22:02AM +0800, xf2017140389@gmail.com wrote:
> From: xiaofeng <xiaofeng5@xiaomi.com>
> 
> When the process_madvise() system call is being used, it takes mmap_lock and blocks the application from allocating memory, leading to unreasonable delays. This patchset aims to fix it.

This means that any page fault will cause madvise() to fail.  NACK.
diff mbox series

Patch

diff --git a/mm/madvise.c b/mm/madvise.c
index 8c927202bbe6..8f1e325873e0 100644
--- a/mm/madvise.c
+++ b/mm/madvise.c
@@ -322,6 +322,9 @@  static int madvise_cold_or_pageout_pte_range(pmd_t *pmd,
 	if (fatal_signal_pending(current))
 		return -EINTR;
 
+	if (mmap_lock_is_contended(mm))
+		return -EINTR;
+
 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 	if (pmd_trans_huge(*pmd)) {
 		pmd_t orig_pmd;