Re: [PATCH] relay: handle alloc_percpu returning NULL in relay_open

[Michael Ellerman]

Daniel Axtens <dja@xxxxxxxxxx> writes:
> alloc_percpu() may return NULL, which means chan->buf may be set to
> NULL. In that case, when we do *per_cpu_ptr(chan->buf, ...), we
> dereference an invalid pointer:
>
> BUG: Unable to handle kernel data access at 0x7dae0000
> Faulting instruction address: 0xc0000000003f3fec
> ...
> NIP [c0000000003f3fec] relay_open+0x29c/0x600
> LR [c0000000003f3fc0] relay_open+0x270/0x600
> Call Trace:
> [c000000054353a70] [c0000000003f3fb4] relay_open+0x264/0x600 (unreliable)
> [c000000054353b00] [c000000000451764] __blk_trace_setup+0x254/0x600
> [c000000054353bb0] [c000000000451b78] blk_trace_setup+0x68/0xa0
> [c000000054353c10] [c0000000010da77c] sg_ioctl+0x7bc/0x2e80
> [c000000054353cd0] [c000000000758cbc] do_vfs_ioctl+0x13c/0x1300
> [c000000054353d90] [c000000000759f14] ksys_ioctl+0x94/0x130
> [c000000054353de0] [c000000000759ff8] sys_ioctl+0x48/0xb0
> [c000000054353e20] [c00000000000bcd0] system_call+0x5c/0x68
>
> Check if alloc_percpu returns NULL. Because we can readily catch and
> handle this situation, switch to alloc_cpu_gfp and pass in __GFP_NOWARN.
>
> This was found by syzkaller both on x86 and powerpc, and the reproducer
> it found on powerpc is capable of hitting the issue as an unprivileged
> user.
>
> Fixes: 017c59c042d0 ("relay: Use per CPU constructs for the relay channel buffer pointers")
> Reported-by: syzbot+1e925b4b836afe85a1c6@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
> Reported-by: syzbot+587b2421926808309d21@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
> Reported-by: syzbot+58320b7171734bf79d26@xxxxxxxxxxxxxxxxxxxxxxxxx
> Reported-by: syzbot+d6074fb08bdb2e010520@xxxxxxxxxxxxxxxxxxxxxxxxx
> Cc: Akash Goel <akash.goel@xxxxxxxxx>
> Cc: Andrew Donnellan <ajd@xxxxxxxxxxxxx> # syzkaller-ppc64
> Cc: stable@xxxxxxxxxxxxxxx # v4.10+
> Signed-off-by: Daniel Axtens <dja@xxxxxxxxxx>
...
> diff --git a/kernel/relay.c b/kernel/relay.c
> index ade14fb7ce2e..a376cc6b54ec 100644
> --- a/kernel/relay.c
> +++ b/kernel/relay.c
> @@ -580,7 +580,13 @@ struct rchan *relay_open(const char *base_filename,
>  	if (!chan)
>  		return NULL;
>  
> -	chan->buf = alloc_percpu(struct rchan_buf *);
> +	chan->buf = alloc_percpu_gfp(struct rchan_buf *,
> +				     GFP_KERNEL | __GFP_NOWARN);
> +	if (!chan->buf) {
> +		kfree(chan);
> +		return NULL;
> +	}
> +
>  	chan->version = RELAYFS_CHANNEL_VERSION;
>  	chan->n_subbufs = n_subbufs;
>  	chan->subbuf_size = subbuf_size;

This looks right to me. The kfree + direct return is correct, there's
nothing else that needs tear down in this function.

I think I'm 50/50 on the __GFP_NOWARN. We're only asking for 8 bytes per
cpu, and if that fails the system is pretty sick, so a warning could be
helpful. There's also logic in the percpu allocator to limit the number
of warnings printed. But see what others think.

Reviewed-by: Michael Ellerman <mpe@xxxxxxxxxxxxxx>

cheers