Date:   Mon, 30 Sep 2019 10:29:53 -0400
From:   Masayoshi Mizuma <msys.mizuma@...il.com>
To:     Dave Martin <Dave.Martin@....com>,
        Julien Grall <julien.grall@....com>
Cc:     Catalin Marinas <catalin.marinas@....com>,
        Will Deacon <will@...nel.org>,
        linux-arm-kernel@...ts.infradead.org,
        Masayoshi Mizuma <m.mizuma@...fujitsu.com>,
        linux-kernel@...r.kernel.org
Subject: Re: [PATCH v2] arm64/sve: Fix wrong free for task->thread.sve_state

Hi Julien and Dave,

On Mon, Sep 30, 2019 at 02:02:46PM +0100, Dave Martin wrote:
> On Mon, Sep 30, 2019 at 01:23:18PM +0100, Julien Grall wrote:
> > Hi,
> > 
> > On 27/09/2019 16:39, Masayoshi Mizuma wrote:
> > >From: Masayoshi Mizuma <m.mizuma@...fujitsu.com>
> > >
> > >The system which has SVE feature crashed because of
> > >the memory pointed by task->thread.sve_state was destroyed
> > >by someone.
> > >
> > >That is because sve_state is freed while the forking the
> > >child process. The child process has the pointer of sve_state
> > >which is same as the parent's because the child's task_struct
> > >is copied from the parent's one. If the copy_process()
> > >fails as an error on somewhere, for example, copy_creds(),
> > >then the sve_state is freed even if the parent is alive.
> > >The flow is as follows.
> > >
> > >copy_process
> > >         p = dup_task_struct
> > >             => arch_dup_task_struct
> > >                 *dst = *src;  // copy the entire region.
> > >:
> > >         retval = copy_creds
> > >         if (retval < 0)
> > >                 goto bad_fork_free;
> > >:
> > >bad_fork_free:
> > >...
> > >         delayed_free_task(p);
> > >           => free_task
> > >              => arch_release_task_struct
> > >                 => fpsimd_release_task
> > >                    => __sve_free
> > >                       => kfree(task->thread.sve_state);
> > >                          // free the parent's sve_state
> > >
> > >Move child's sve_state = NULL and clearing TIF_SVE flag
> > >to arch_dup_task_struct() so that the child doesn't free the
> > >parent's one.
> > >
> > >Cc: stable@...r.kernel.org
> > >Fixes: bc0ee4760364 ("arm64/sve: Core task context handling")
> > 
> > Looking at the log, it looks like THREAD_INFO_IN_TASK was selected before
> > the bc0ee4760364. So it should be fine to backport for all the Linux tree
> > contain this commit.

I think this patch is needed for the kernel has SVE support.
I'll add the Cc tag as Dave said:

Cc: stable@...r.kernel.org # 4.15+

So, I suppose this patch will be backported to stables 5.3.X,
5.2.X and longterm 4.19.X.
Does this make sense?

> > 
> > >Signed-off-by: Masayoshi Mizuma <m.mizuma@...fujitsu.com>
> > >Reported-by: Hidetoshi Seto <seto.hidetoshi@...fujitsu.com>
> > >Suggested-by: Dave Martin <Dave.Martin@....com>
> > 
> > I have tested the patch and can confirm that double-free disappeared after
> > the patch is applied:
> > 
> > Tested-by: Julien Grall <julien.grall@....com>

Thank you so much!

> 
> Good to have that confirmed -- thanks for verifying.
> 
> [...]
> 
> > >---
> > >  arch/arm64/kernel/process.c | 21 ++++-----------------
> > >  1 file changed, 4 insertions(+), 17 deletions(-)
> > >
> > >diff --git a/arch/arm64/kernel/process.c b/arch/arm64/kernel/process.c
> > >index f674f28df..6937f5935 100644
> > >--- a/arch/arm64/kernel/process.c
> > >+++ b/arch/arm64/kernel/process.c
> > >@@ -323,22 +323,16 @@ void arch_release_task_struct(struct task_struct *tsk)
> > >  	fpsimd_release_task(tsk);
> > >  }
> > >-/*
> > >- * src and dst may temporarily have aliased sve_state after task_struct
> > >- * is copied.  We cannot fix this properly here, because src may have
> > >- * live SVE state and dst's thread_info may not exist yet, so tweaking
> > >- * either src's or dst's TIF_SVE is not safe.
> > >- *
> > >- * The unaliasing is done in copy_thread() instead.  This works because
> > >- * dst is not schedulable or traceable until both of these functions
> > >- * have been called.
> > >- */
> > 
> > It would be good to explain in the commit message why tweaking "dst" in
> > arch_dup_task_struct() is fine.
> > 
> > From my understanding, Arm64 used to have thread_info on the stack. So it
> > would not be possible to clear TIF_SVE until the stack is initialized.
> > 
> > Now that the thread_info is part of the task, it should be valid to modify
> > the flag from arch_dup_task_struct().
> > 
> > Note that technically, TIF_SVE does not need to be cleared from
> > arch_dup_task_struct(). It could also be done from copy_thread(). But it is
> > easier to keep the both changes together.

Thanks, let me add some comments to the commit log.

> > 
> > >  int arch_dup_task_struct(struct task_struct *dst, struct task_struct *src)
> > >  {
> > >  	if (current->mm)
> > >  		fpsimd_preserve_current_state();
> > >  	*dst = *src;
> 
> Ack, some more explanation would be a good idea here.
> 
> Maybe the following comments are sufficient?
> 
> 	/* We rely on the above assingment to initialise dst's thread_flags: */

Thanks, I'll add this comment.

> 
> > >+	BUILD_BUG_ON(!IS_ENABLED(CONFIG_THREAD_INFO_IN_TASK));
> > 
> 
> and
> 
> 	/*
> 	 * Detach src's sve_state (if any) from dst so that it does not
> 	 * get erroneously used or freed prematurely.  dst's sve_state
> 	 * will be allocated on demand later on if dst uses SVE.
> 	 * For consistency, also clear TIF_SVE here: this could be done
> 	 * later in copy_process(), but to avoid tripping up future
> 	 * maintainers it is best not to leave TIF_SVE and sve_state in
> 	 * an inconsistent state, even temporarily.
> 	 */

I'll add this comments.

> 
> > >+	dst->thread.sve_state = NULL;
> > >+	clear_tsk_thread_flag(dst, TIF_SVE);
> 
> (TIF_SVE should not usually be set in the first place of course, since
> we are in a fork() or clone() syscall in src.  This may not be true if
> a task is created using kernel_thread() while running in the context of
> some user task that entered the kernel due to a trap or syscall --
> though probably nobody should be doing that.)

Thanks!
Masa

Powered by blists - more mailing lists