Re: [PATCH] mm: make munlock fast when mlock is canceled by sigkill

[Minchan Kim <minchan.kim@gmail.com>]

On Sun, Aug 23, 2009 at 1:54 AM, Hiroaki
Wakabayashi<primulaelatior@gmail.com> wrote:
> From 27b2fde0222c59049026e7d0bdc4a2a68d0720f5 Mon Sep 17 00:00:00 2001
> From: Hiroaki Wakabayashi <primulaelatior@gmail.com>
> Date: Sat, 22 Aug 2009 19:14:53 +0900
> Subject: [PATCH] mm: make munlock fast when mlock is canceled by sigkill
>
> This patch is for making commit 4779280d1e (mm: make get_user_pages()
> interruptible) complete.
>
> At first, munlock() assumes that all pages in vma are pinned,
>
> Now, by the commit, mlock() can be interrupted by SIGKILL, etc  So, part of
> pages are not pinned.
> If SIGKILL, In exit() path, munlock is called for unlocking pinned pages
> in vma.
>
> But, there, get_user_pages(write) is used for munlock(). Then, pages are
> allocated via page-fault for exsiting process !!! This is problem at canceling
> big mlock.
> This patch tries to avoid allocating new pages at munlock().
>
>   mlock( big area )
>        <===== sig kill
>   do_exit()
>    ->mmput()
>       -> do_munlock()
>         -> get_user_pages()
>               <allocate *never used* memory>
>       ->.....freeing allocated memory.
>
> * Test program
> % cat run.sh
> #!/bin/sh
>
> ./mlock_test 2000000000 &
> sleep 2
> kill -9 $!
> wait
>
> % cat mlock_test.c
> #include <stdio.h>
> #include <stdlib.h>
> #include <string.h>
> #include <sys/mman.h>
> #include <sys/types.h>
> #include <sys/stat.h>
> #include <fcntl.h>
> #include <errno.h>
> #include <time.h>
> #include <unistd.h>
> #include <sys/time.h>
>
> int main(int argc, char **argv)
> {
>        size_t length = 50 * 1024 * 1024;
>        void *addr;
>        time_t timer;
>
>        if (argc >= 2)
>                length = strtoul(argv[1], NULL, 10);
>        printf("PID = %d\n", getpid());
>        addr = mmap(NULL, length, PROT_READ | PROT_WRITE,
>                                MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
>        if (addr == MAP_FAILED) {
>                fprintf(stderr, "mmap failed: %s, length=%lu\n",
>                                strerror(errno), length);
>                exit(EXIT_FAILURE);
>        }
>        printf("try mlock length=%lu\n", length);
>        timer = time(NULL);
>        if (mlock(addr, length) < 0) {
>                fprintf(stderr, "mlock failed: %s, time=%lu[sec]\n",
>                                strerror(errno), time(NULL) - timer);
>                exit(EXIT_FAILURE);
>        }
>        printf("mlock succeed, time=%lu[sec]\n\n", time(NULL) - timer);
>        printf("try munlock length=%lu\n", length);
>        timer = time(NULL);
>        if (munlock(addr, length) < 0) {
>                fprintf(stderr, "munlock failed: %s, time=%lu[sec]\n",
>                                strerror(errno), time(NULL)-timer);
>                exit(EXIT_FAILURE);
>        }
>        printf("munlock succeed, time=%lu[sec]\n\n", time(NULL) - timer);
>        if (munmap(addr, length) < 0) {
>                fprintf(stderr, "munmap failed: %s\n", strerror(errno));
>                exit(EXIT_FAILURE);
>        }
>        return 0;
> }
>
> * Executed Result
> -- Original executed result
> % time ./run.sh
>
> PID = 2678
> try mlock length=2000000000
> ./run.sh: line 6:  2678 Killed                  ./mlock_test 2000000000
> ./run.sh  0.00s user 2.59s system 13% cpu 18.781 total
> %
>
> -- After applied this patch
> % time ./run.sh
>
> PID = 2512
> try mlock length=2000000000
> ./run.sh: line 6:  2512 Killed                  ./mlock_test 2000000000
> ./run.sh  0.00s user 1.15s system 45% cpu 2.507 total
> %
>
> Signed-off-by: Hiroaki Wakabayashi <primulaelatior@gmail.com>
> ---
>  mm/internal.h |    1 +
>  mm/memory.c   |    9 +++++++--
>  mm/mlock.c    |   35 +++++++++++++++++++----------------
>  3 files changed, 27 insertions(+), 18 deletions(-)
>
> diff --git a/mm/internal.h b/mm/internal.h
> index f290c4d..4ab5b24 100644
> --- a/mm/internal.h
> +++ b/mm/internal.h
> @@ -254,6 +254,7 @@ static inline void
> mminit_validate_memmodel_limits(unsigned long *start_pfn,
>  #define GUP_FLAGS_FORCE                  0x2
>  #define GUP_FLAGS_IGNORE_VMA_PERMISSIONS 0x4
>  #define GUP_FLAGS_IGNORE_SIGKILL         0x8
> +#define GUP_FLAGS_ALLOW_NULL             0x10
>

I am worried about adding new flag whenever we need it.
But I think this case makes sense to me.
In addition, I guess ZERO page can also use this flag.

Kame. What do you think about it?


>  int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
>                     unsigned long start, int len, int flags,
> diff --git a/mm/memory.c b/mm/memory.c
> index aede2ce..b41fbf9 100644
> --- a/mm/memory.c
> +++ b/mm/memory.c
> @@ -1217,6 +1217,7 @@ int __get_user_pages(struct task_struct *tsk,
> struct mm_struct *mm,
>        int force = !!(flags & GUP_FLAGS_FORCE);
>        int ignore = !!(flags & GUP_FLAGS_IGNORE_VMA_PERMISSIONS);
>        int ignore_sigkill = !!(flags & GUP_FLAGS_IGNORE_SIGKILL);
> +       int allow_null = !!(flags & GUP_FLAGS_ALLOW_NULL);
>
>        if (nr_pages <= 0)
>                return 0;
> @@ -1312,6 +1313,8 @@ int __get_user_pages(struct task_struct *tsk,
> struct mm_struct *mm,
>                        while (!(page = follow_page(vma, start, foll_flags))) {
>                                int ret;
>
> +                               if (allow_null)
> +                                       break;
>                                ret = handle_mm_fault(mm, vma, start,
>                                        (foll_flags & FOLL_WRITE) ?
>                                        FAULT_FLAG_WRITE : 0);
> @@ -1351,8 +1354,10 @@ int __get_user_pages(struct task_struct *tsk,
> struct mm_struct *mm,
>                        if (pages) {
>                                pages[i] = page;
>
> -                               flush_anon_page(vma, page, start);
> -                               flush_dcache_page(page);
> +                               if (page) {
> +                                       flush_anon_page(vma, page, start);
> +                                       flush_dcache_page(page);
> +                               }
>                        }
>                        if (vmas)
>                                vmas[i] = vma;
> diff --git a/mm/mlock.c b/mm/mlock.c
> index 45eb650..0f5827b 100644
> --- a/mm/mlock.c
> +++ b/mm/mlock.c
> @@ -178,9 +178,10 @@ static long __mlock_vma_pages_range(struct
> vm_area_struct *vma,
>         */
>        if (!mlock)
>                gup_flags |= GUP_FLAGS_IGNORE_VMA_PERMISSIONS |
> -                            GUP_FLAGS_IGNORE_SIGKILL;
> +                            GUP_FLAGS_IGNORE_SIGKILL |
> +                            GUP_FLAGS_ALLOW_NULL;
>
> -       if (vma->vm_flags & VM_WRITE)
> +       if (mlock && (vma->vm_flags & VM_WRITE))
>                gup_flags |= GUP_FLAGS_WRITE;
>
>        while (nr_pages > 0) {
> @@ -220,21 +221,23 @@ static long __mlock_vma_pages_range(struct
> vm_area_struct *vma,
>                for (i = 0; i < ret; i++) {
>                        struct page *page = pages[i];
>
> -                       lock_page(page);
> -                       /*
> -                        * Because we lock page here and migration is blocked
> -                        * by the elevated reference, we need only check for
> -                        * page truncation (file-cache only).
> -                        */
> -                       if (page->mapping) {
> -                               if (mlock)
> -                                       mlock_vma_page(page);
> -                               else
> -                                       munlock_vma_page(page);
> +                       if (page) {
> +                               lock_page(page);
> +                               /*
> +                                * Because we lock page here and migration is
> +                                * blocked by the elevated reference, we need
> +                                * only check for page truncation
> +                                * (file-cache only).
> +                                */
> +                               if (page->mapping) {
> +                                       if (mlock)
> +                                               mlock_vma_page(page);
> +                                       else
> +                                               munlock_vma_page(page);
> +                               }
> +                               unlock_page(page);
> +                               put_page(page); /* ref from get_user_pages() */
>                        }
> -                       unlock_page(page);
> -                       put_page(page);         /* ref from get_user_pages() */
> -
>                        /*
>                         * here we assume that get_user_pages() has given us
>                         * a list of virtually contiguous pages.
> --
> 1.5.6.5
> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at  http://www.tux.org/lkml/
>



-- 
Kind regards,
Minchan Kim