Re: [PATCH v3] Support madvise(MADV_DONTDUMP) when creating core dumps for qemu-user

[Daniel P . Berrangé]

On Tue, May 06, 2025 at 05:46:02PM +0100, WorksButNotTested wrote:
> When running applications which make large (sparsely populated) address ranges
> (e.g. when using address sanitizer with LibAFL) the inability to exclude these
> regions from any core dump can result in very large files which fill the disk.
> A coredump is obvously very useful for performing a post-mortem when fuzzing.
> 
> Whilst the man pages state that madvise provides only a hint (and hence can be
> ignored), this patch adds support to handle MADV_DONTDUMP and set a
> corresponding flag in the page flags, thus allowing QEMU to exclude these
> regions from the core file.
> 
> Signed-off-by: WorksButNotTested 
> <62701594+WorksButNotTested@users.noreply.github.com>

Any reason you've not used your "jonwilson030981@googlemail.com"
address for this.

This github alias rejects any mail delivery, so also should not
be CC'd on the patch either, as that triggers failures when
reviewers reply to this submission.

> ---
>  include/exec/page-protection.h |  6 ++++++
>  linux-user/elfload.c           |  4 ++++
>  linux-user/mmap.c              | 18 ++++++++++++++++++
>  3 files changed, 28 insertions(+)
> 
> diff --git a/include/exec/page-protection.h b/include/exec/page-protection.h
> index c43231af8b..f8826d917e 100644
> --- a/include/exec/page-protection.h
> +++ b/include/exec/page-protection.h
> @@ -38,4 +38,10 @@
>   */
>  #define PAGE_PASSTHROUGH 0x0800
>  
> +/*
> + * For linux-user, indicates that the page should not be included in a core 
> + * dump.
> + */
> +#define PAGE_DONTDUMP   0x1000
> +
>  #endif
> diff --git a/linux-user/elfload.c b/linux-user/elfload.c
> index fbfdec2f17..41c46da055 100644
> --- a/linux-user/elfload.c
> +++ b/linux-user/elfload.c
> @@ -4067,6 +4067,10 @@ static size_t vma_dump_size(target_ulong start, 
> target_ulong end,
>          return 0;
>      }
>  
> +    if (flags & PAGE_DONTDUMP) {
> +        return 0;
> +    }
> +
>      /*
>       * Usually we don't dump executable pages as they contain
>       * non-writable code that debugger can read directly from
> diff --git a/linux-user/mmap.c b/linux-user/mmap.c
> index f88a80c31e..016063a8cf 100644
> --- a/linux-user/mmap.c
> +++ b/linux-user/mmap.c
> @@ -1247,6 +1247,24 @@ abi_long target_madvise(abi_ulong start, abi_ulong 
> len_in, int advice)
>       */
>      mmap_lock();
>      switch (advice) {
> +    case MADV_DONTDUMP:
> +        if (len > 0) {
> +            /*
> +             * To set the page permissons, we must OR our new flags with the
> +             * existing flags. Only mark the pages as PAGE_DONTDUMP if the
> +             * entire range has the same flags. If any part of the range
> +             * differs, we would need to process it one page at a time which
> +             * might not be very performant. Since we are not obliged to 
> respect
> +             * this flag, we will support it for the most likely usage 
> scenario.
> +             * Note that we don't set PAGE_ANON, since this can only be set 
> with
> +             * new mappings.
> +             */
> +            int flg = page_get_flags(start);
> +            if (page_check_range(start, len, flg)) {
> +                page_set_flags(start, start + len - 1, PAGE_DONTDUMP | (flg 
> & ~PAGE_ANON) );
> +            }
> +        }
> +        break;
>      case MADV_WIPEONFORK:
>      case MADV_KEEPONFORK:
>          ret = -EINVAL;
> -- 
> 2.43.0
> 
> 

With regards,
Daniel
-- 
|: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
|: https://libvirt.org         -o-            https://fstop138.berrange.com :|
|: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|