Re: [PATCH] target/user: Fix use-after-free cmd->se_cmd if the cmd isexpired

[Xiubo Li]

Hi Mike

Thanks very much for your analysis.

> > diff --git a/drivers/target/target_core_user.c b/drivers/target/target_core_user.c
> > index 2e33100..6396581 100644
> > --- a/drivers/target/target_core_user.c
> > +++ b/drivers/target/target_core_user.c
> > @@ -684,7 +684,6 @@ static int tcmu_check_expired_cmd(int id, void *p, void *data)
> >   
> >   	set_bit(TCMU_CMD_BIT_EXPIRED, &cmd->flags);
> >   	target_complete_cmd(cmd->se_cmd, SAM_STAT_CHECK_CONDITION);
> > -	cmd->se_cmd = NULL;
> >   
> How did tcmu_handle_completion get to a point it was accessing the
> se_cmd if the TCMU_CMD_BIT_EXPIRED bit was set?
> Were memory accesses out
> of order?
No, even using the -O3, becuase has there memory dependency ?

> CPU1 set the TCMU_CMD_BIT_EXPIRED bit then cleared
> cmd->se_cmd, but CPU2 copied cmd->se_cmd to se_cmd and saw it was NULL
> but did not yet see the TCMU_CMD_BIT_EXPIRED bit set?

Because the debug rpms for my kernel version were lost, and the crash
tools couldn't be used to have a more accurate analysis.
> It looks like, if you do the above patch, the above function will call
> target_complete_cmd and tcmu_handle_completion will call it again, so we
> will have a double free issue.
Maybe the best resolution is to move tcmu_handle_completion() between
spin_lock(&udev->commands_lock) and spin_unlock(&udev->commands_lock)?

Thanks.

BRs
Xiubo Li