Hi Eric,
On Sun, Jun 14, 2020 at 10:00:19PM -0700, Eric Biggers wrote:
> On Mon, Jun 15, 2020 at 09:29:48AM +0530, Sahitya Tummala wrote:
> > There could be a potential race between these two paths below,
> > leading to use-after-free when accessing bio->bi_crypt_context.
> >
> > f2fs_write_cache_pages
> > ->f2fs_do_write_data_page on page#1
> > ->f2fs_inplace_write_data
> > ->f2fs_merge_page_bio
> > ->add_bio_entry
> > ->f2fs_do_write_data_page on page#2
> > ->f2fs_inplace_write_data
> > ->f2fs_merge_page_bio
> > ->f2fs_crypt_mergeable_bio
> > ->fscrypt_mergeable_bio
> > f2fs_write_begin on page#1
> > ->f2fs_wait_on_page_writeback
> > ->f2fs_submit_merged_ipu_write
> > ->__submit_bio
> > The bio gets completed, calling
> > bio_endio
> > ->bio_uninit
> > ->bio_crypt_free_ctx
> > ->use-after-free issue
> >
> > Fix this by moving f2fs_crypt_mergeable_bio() check within
> > add_ipu_page() so that it's done under bio_list_lock to prevent
> > the above race.
> >
> > Fixes: 15e76ad23e72 ("f2fs: add inline encryption support")
> > Signed-off-by: Sahitya Tummala <stumm...@codeaurora.org>
> > ---
> > This fix is rebased to the tip of fscrypt git -
> > https://git.kernel.org/pub/scm/fs/fscrypt/fscrypt.git
> > branch - inline-encryption
> >
> > fs/f2fs/data.c | 26 ++++++++++++++++++--------
> > 1 file changed, 18 insertions(+), 8 deletions(-)
> >
> > diff --git a/fs/f2fs/data.c b/fs/f2fs/data.c
> > index 0dfa8d3..3b53554 100644
> > --- a/fs/f2fs/data.c
> > +++ b/fs/f2fs/data.c
> > @@ -762,9 +762,10 @@ static void del_bio_entry(struct bio_entry *be)
> > kmem_cache_free(bio_entry_slab, be);
> > }
> >
> > -static int add_ipu_page(struct f2fs_sb_info *sbi, struct bio **bio,
> > - struct page *page)
> > +static int add_ipu_page(struct f2fs_io_info *fio, struct bio **bio,
> > + struct page *page, int *bio_needs_submit)
> > {
> > + struct f2fs_sb_info *sbi = fio->sbi;
> > enum temp_type temp;
> > bool found = false;
> > int ret = -EAGAIN;
> > @@ -780,6 +781,15 @@ static int add_ipu_page(struct f2fs_sb_info *sbi,
> > struct bio **bio,
> > continue;
> >
> > found = true;
> > + if (*bio && (!page_is_mergeable(sbi, *bio,
> > + *fio->last_block, fio->new_blkaddr) ||
> > + !f2fs_crypt_mergeable_bio(*bio,
> > + fio->page->mapping->host,
> > + fio->page->index, fio))) {
> > + ret = 0;
> > + *bio_needs_submit = 1;
> > + break;
> > + }
> >
> > if (bio_add_page(*bio, page, PAGE_SIZE, 0) ==
> > PAGE_SIZE) {
> > @@ -864,6 +874,7 @@ int f2fs_merge_page_bio(struct f2fs_io_info *fio)
> > struct bio *bio = *fio->bio;
> > struct page *page = fio->encrypted_page ?
> > fio->encrypted_page : fio->page;
> > + int bio_needs_submit = 0;
> >
> > if (!f2fs_is_valid_blkaddr(fio->sbi, fio->new_blkaddr,
> > __is_meta_io(fio) ? META_GENERIC : DATA_GENERIC))
> > @@ -872,11 +883,6 @@ int f2fs_merge_page_bio(struct f2fs_io_info *fio)
> > trace_f2fs_submit_page_bio(page, fio);
> > f2fs_trace_ios(fio, 0);
> >
> > - if (bio && (!page_is_mergeable(fio->sbi, bio, *fio->last_block,
> > - fio->new_blkaddr) ||
> > - !f2fs_crypt_mergeable_bio(bio, fio->page->mapping->host,
> > - fio->page->index, fio)))
> > - f2fs_submit_merged_ipu_write(fio->sbi, &bio, NULL);
> > alloc_new:
> > if (!bio) {
> > bio = __bio_alloc(fio, BIO_MAX_PAGES);
> > @@ -886,8 +892,12 @@ int f2fs_merge_page_bio(struct f2fs_io_info *fio)
> >
> > add_bio_entry(fio->sbi, bio, page, fio->temp);
> > } else {
> > - if (add_ipu_page(fio->sbi, &bio, page))
> > + if (add_ipu_page(fio, &bio, page, &bio_needs_submit))
> > + goto alloc_new;
> > + if (bio_needs_submit) {
> > + f2fs_submit_merged_ipu_write(fio->sbi, &bio, NULL);
> > goto alloc_new;
> > + }
> > }
> >
> > if (fio->io_wbc)
>
> Thanks, I'm still trying to understand this part of the code, but it's looking
> like this is a real bug. Do you also have a reproducer that produces a KASAN
> report, or did you find this another way?
>
This was once reported as a stability issue internally on our downstream kernel
having the new FBE patches. I have then reviewed the code and found the issue
path.
There is no consistent way to reproduce this problem.
> One comment: add_ipu_page() already submits the bio if it's full. Wouldn't it
> be better to use that instead of f2fs_submit_merged_ipu_write()? I.e.:
>
Your suggestion looks good to me. I will post a v2.
Thanks,
> diff --git a/fs/f2fs/data.c b/fs/f2fs/data.c
> index e9dcda80e599..d7a51dbe208b 100644
> --- a/fs/f2fs/data.c
> +++ b/fs/f2fs/data.c
> @@ -762,9 +762,10 @@ static void del_bio_entry(struct bio_entry *be)
> kmem_cache_free(bio_entry_slab, be);
> }
>
> -static int add_ipu_page(struct f2fs_sb_info *sbi, struct bio **bio,
> +static int add_ipu_page(struct f2fs_io_info *fio, struct bio **bio,
> struct page *page)
> {
> + struct f2fs_sb_info *sbi = fio->sbi;
> enum temp_type temp;
> bool found = false;
> int ret = -EAGAIN;
> @@ -780,14 +781,18 @@ static int add_ipu_page(struct f2fs_sb_info *sbi,
> struct bio **bio,
> continue;
>
> found = true;
> -
> - if (bio_add_page(*bio, page, PAGE_SIZE, 0) ==
> - PAGE_SIZE) {
> + if (page_is_mergeable(sbi, *bio, *fio->last_block,
> + fio->new_blkaddr) &&
> + f2fs_crypt_mergeable_bio(*bio,
> + fio->page->mapping->host,
> + fio->page->index, fio) &&
> + bio_add_page(*bio, page,
> + PAGE_SIZE, 0) == PAGE_SIZE) {
> ret = 0;
> break;
> }
>
> - /* bio is full */
> + /* page can't be merged into bio; submit the bio */
> del_bio_entry(be);
> __submit_bio(sbi, *bio, DATA);
> break;
> @@ -872,11 +877,6 @@ int f2fs_merge_page_bio(struct f2fs_io_info *fio)
> trace_f2fs_submit_page_bio(page, fio);
> f2fs_trace_ios(fio, 0);
>
> - if (bio && (!page_is_mergeable(fio->sbi, bio, *fio->last_block,
> - fio->new_blkaddr) ||
> - !f2fs_crypt_mergeable_bio(bio, fio->page->mapping->host,
> - fio->page->index, fio)))
> - f2fs_submit_merged_ipu_write(fio->sbi, &bio, NULL);
> alloc_new:
> if (!bio) {
> bio = __bio_alloc(fio, BIO_MAX_PAGES);
> @@ -886,7 +886,7 @@ int f2fs_merge_page_bio(struct f2fs_io_info *fio)
>
> add_bio_entry(fio->sbi, bio, page, fio->temp);
> } else {
> - if (add_ipu_page(fio->sbi, &bio, page))
> + if (add_ipu_page(fio, &bio, page))
> goto alloc_new;
> }
>
--
--
Sent by a consultant of the Qualcomm Innovation Center, Inc.
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum.
