Re: [PATCH v4 1/2] fork: extend clone3() to support CLONE_SET_TID

From: Christian Brauner
Date: Sun Aug 11 2019 - 15:08:46 EST


On Sun, Aug 11, 2019 at 08:51:48AM +0200, Christian Brauner wrote:
> On Sat, Aug 10, 2019 at 07:59:18AM +0200, Adrian Reber wrote:
> > On Sat, Aug 10, 2019 at 03:10:34AM +0200, Christian Brauner wrote:
> > > On Thu, Aug 08, 2019 at 11:22:21PM +0200, Adrian Reber wrote:
> > > > The main motivation to add set_tid to clone3() is CRIU.
> > > >
> > > > To restore a process with the same PID/TID CRIU currently uses
> > > > /proc/sys/kernel/ns_last_pid. It writes the desired (PID - 1) to
> > > > ns_last_pid and then (quickly) does a clone(). This works most of the
> > > > time, but it is racy. It is also slow as it requires multiple syscalls.
> > > >
> > > > Extending clone3() to support set_tid makes it possible restore a
> > > > process using CRIU without accessing /proc/sys/kernel/ns_last_pid and
> > > > race free (as long as the desired PID/TID is available).
> > > >
> > > > This clone3() extension places the same restrictions (CAP_SYS_ADMIN)
> > > > on clone3() with set_tid as they are currently in place for ns_last_pid.
> > > >
> > > > Signed-off-by: Adrian Reber <areber@xxxxxxxxxx>
> > > > ---
> > > > v2:
> > > > - Removed (size < sizeof(struct clone_args)) as discussed with
> > > > Christian and Dmitry
> > > > - Added comment to ((set_tid != 1) && idr_get_cursor() <= 1) (Oleg)
> > > > - Use idr_alloc() instead of idr_alloc_cyclic() (Oleg)
> > > >
> > > > v3:
> > > > - Return EEXIST if PID is already in use (Christian)
> > > > - Drop CLONE_SET_TID (Christian and Oleg)
> > > > - Use idr_is_empty() instead of idr_get_cursor() (Oleg)
> > > > - Handle different `struct clone_args` sizes (Dmitry)
> > > >
> > > > v4:
> > > > - Rework struct size check with defines (Christian)
> > > > - Reduce number of set_tid checks (Oleg)
> > > > - Less parentheses and more robust code (Oleg)
> > > > - Do ns_capable() on correct user_ns (Oleg, Christian)
> > > > ---
> > > > include/linux/pid.h | 2 +-
> > > > include/linux/sched/task.h | 1 +
> > > > include/uapi/linux/sched.h | 1 +
> > > > kernel/fork.c | 25 +++++++++++++++++++++++--
> > > > kernel/pid.c | 34 +++++++++++++++++++++++++++-------
> > > > 5 files changed, 53 insertions(+), 10 deletions(-)
> > > >
> > > > diff --git a/include/linux/pid.h b/include/linux/pid.h
> > > > index 2a83e434db9d..052000db0ced 100644
> > > > --- a/include/linux/pid.h
> > > > +++ b/include/linux/pid.h
> > > > @@ -116,7 +116,7 @@ extern struct pid *find_vpid(int nr);
> > > > extern struct pid *find_get_pid(int nr);
> > > > extern struct pid *find_ge_pid(int nr, struct pid_namespace *);
> > > >
> > > > -extern struct pid *alloc_pid(struct pid_namespace *ns);
> > > > +extern struct pid *alloc_pid(struct pid_namespace *ns, pid_t set_tid);
> > > > extern void free_pid(struct pid *pid);
> > > > extern void disable_pid_allocation(struct pid_namespace *ns);
> > > >
> > > > diff --git a/include/linux/sched/task.h b/include/linux/sched/task.h
> > > > index 0497091e40c1..4f2a80564332 100644
> > > > --- a/include/linux/sched/task.h
> > > > +++ b/include/linux/sched/task.h
> > > > @@ -26,6 +26,7 @@ struct kernel_clone_args {
> > > > unsigned long stack;
> > > > unsigned long stack_size;
> > > > unsigned long tls;
> > > > + pid_t set_tid;
> > > > };
> > > >
> > > > /*
> > > > diff --git a/include/uapi/linux/sched.h b/include/uapi/linux/sched.h
> > > > index b3105ac1381a..e1ce103a2c47 100644
> > > > --- a/include/uapi/linux/sched.h
> > > > +++ b/include/uapi/linux/sched.h
> > > > @@ -45,6 +45,7 @@ struct clone_args {
> > > > __aligned_u64 stack;
> > > > __aligned_u64 stack_size;
> > > > __aligned_u64 tls;
> > > > + __aligned_u64 set_tid;
> > > > };
> > > >
> > > > /*
> > > > diff --git a/kernel/fork.c b/kernel/fork.c
> > > > index 2852d0e76ea3..2a03f0e201e9 100644
> > > > --- a/kernel/fork.c
> > > > +++ b/kernel/fork.c
> > > > @@ -117,6 +117,13 @@
> > > > */
> > > > #define MAX_THREADS FUTEX_TID_MASK
> > > >
> > > > +/*
> > > > + * Different sizes of struct clone_args
> > > > + */
> > > > +#define CLONE3_ARGS_SIZE_V0 64
> > > > +/* V1 includes set_tid */
> > > > +#define CLONE3_ARGS_SIZE_V1 72
> > > > +
> > > > /*
> > > > * Protected counters by write_lock_irq(&tasklist_lock)
> > > > */
> > > > @@ -2031,7 +2038,13 @@ static __latent_entropy struct task_struct *copy_process(
> > > > stackleak_task_init(p);
> > > >
> > > > if (pid != &init_struct_pid) {
> > > > - pid = alloc_pid(p->nsproxy->pid_ns_for_children);
> > > > + if (args->set_tid && !ns_capable(
> > > > + p->nsproxy->pid_ns_for_children->user_ns,
> > > > + CAP_SYS_ADMIN)) {
> > > > + retval = -EPERM;
> > > > + goto bad_fork_cleanup_thread;
> > > > + }
> > > > + pid = alloc_pid(p->nsproxy->pid_ns_for_children, args->set_tid);
> > > > if (IS_ERR(pid)) {
> > > > retval = PTR_ERR(pid);
> > > > goto bad_fork_cleanup_thread;
> > > > @@ -2535,9 +2548,14 @@ noinline static int copy_clone_args_from_user(struct kernel_clone_args *kargs,
> > > > if (unlikely(size > PAGE_SIZE))
> > > > return -E2BIG;
> > > >
> > > > - if (unlikely(size < sizeof(struct clone_args)))
> > > > + /* The struct needs to be at least the size of the original struct. */
> > >
> > > I don't think you need that comment. I think the macro is pretty
> > > self-explanatory. If you want it to be even clearer you could even make
> > > it CLONE3_ARGS_SIZE_MIN but V0 is good enough. :)
> >
> > Will remove the comment.
> >
> > > > + if (unlikely(size < CLONE3_ARGS_SIZE_V0))
> > > > return -EINVAL;
> > > >
> > > > + if (size < sizeof(struct clone_args))
> > > > + memset((void *)&args + size, 0,
> > > > + sizeof(struct clone_args) - size);
> > > > +
> > > > if (unlikely(!access_ok(uargs, size)))
> > > > return -EFAULT;
> > > >
> > > > @@ -2573,6 +2591,9 @@ noinline static int copy_clone_args_from_user(struct kernel_clone_args *kargs,
> > > > .tls = args.tls,
> > > > };
> > > >
> > > > + if (size >= CLONE3_ARGS_SIZE_V1)
> > > > + kargs->set_tid = args.set_tid;
> > >
> > > Hm, the if-condition is not needed though, right? At this point we will
> > > have already copied from struct clone_args __user *uargs into struct
> > > clone_args args. If we hit that codepath that means the kernel
> > > definitely has a field for set_tid in its struct clone_args. :) So this
> > > could probably just be:
> > >
> > > .tls = args.tls,
> > > .set_tid = args.set_tid,
> > > }
> > >
> > > ?
> >
> > Right.
> >
> > > > +
> > > > return 0;
> > > > }
> > > >
> > > > diff --git a/kernel/pid.c b/kernel/pid.c
> > > > index 0a9f2e437217..9ce89c35c5be 100644
> > > > --- a/kernel/pid.c
> > > > +++ b/kernel/pid.c
> > > > @@ -157,7 +157,7 @@ void free_pid(struct pid *pid)
> > > > call_rcu(&pid->rcu, delayed_put_pid);
> > > > }
> > > >
> > > > -struct pid *alloc_pid(struct pid_namespace *ns)
> > > > +struct pid *alloc_pid(struct pid_namespace *ns, int set_tid)
> > > > {
> > > > struct pid *pid;
> > > > enum pid_type type;
> > > > @@ -186,12 +186,32 @@ struct pid *alloc_pid(struct pid_namespace *ns)
> > > > if (idr_get_cursor(&tmp->idr) > RESERVED_PIDS)
> > > > pid_min = RESERVED_PIDS;
> > > >
> > > > - /*
> > > > - * Store a null pointer so find_pid_ns does not find
> > > > - * a partially initialized PID (see below).
> > > > - */
> > > > - nr = idr_alloc_cyclic(&tmp->idr, NULL, pid_min,
> > > > - pid_max, GFP_ATOMIC);
> > > > + if (set_tid) {
> > > > + /*
> > > > + * Also fail if a PID != 1 is requested
> > > > + * and no PID 1 exists.
> > > > + */
> > > > + nr = -EINVAL;
> > > > + if (set_tid < pid_max && set_tid > 0 &&
> > >
> > > Hm, you're already in the if-branch hat verified if (set_tid) so the
> > > set_tid > 0 conjunct seems redundant. :)
> >
> > Yes, but I dropped all checks to see if set_tid is negative as suggested
> > by Oleg and moved it here.
> >
> > > > + (set_tid == 1 || !idr_is_empty(&tmp->idr)))
> > > > + nr = idr_alloc(&tmp->idr, NULL, set_tid,
> > > > + set_tid + 1, GFP_ATOMIC);
> > >
> > > I'm confused, shouldn't this be
> > >
> > > if (set_tid < pid_max || (set_tid == 1 && !idr_is_emtpy(&tmp->idf)))
> >
> > Now I am also confused ;). This does not work. This will always return
> > true if set_tid is less than pid_max. So pid_max needs to be something
> > like 1 for the check after || to make sense, right? But you really got
> > me confused here right now. Right now I still think what I did is
> > correct.
>
> I missed the part where you reset set_tid to 0 below and mis-parsed the
> rightmost conjunct in the if statement.
>
> One thing I dislike is that you do
>
> if (set_tid) {
> if ([...] && set_tid > 0 && [...])
> }

Thinking about this a bit more you probably did the explicit set_tid > 0 to
catch the case where it is negative? But also, I don't understand why we do any
work at all before having verified that set_tid is sensible, i.e. why do we
call kmem_cache_calloc(), do idr_preload(), and take a spinlock, before even
verifying that our parameters are sane? If there's no specific reason for this
I suggest to patch alloc_pid() like this:

-struct pid *alloc_pid(struct pid_namespace *ns)
+struct pid *alloc_pid(struct pid_namespace *ns, int set_tid)
{
struct pid *pid;
enum pid_type type;
@@ -166,6 +166,9 @@ struct pid *alloc_pid(struct pid_namespace *ns)
struct upid *upid;
int retval = -ENOMEM;

+ if (set_tid < 0 || set_tid >= pid_max)
+ return ERR_PTR(-EINVAL);
+
pid = kmem_cache_alloc(ns->pid_cachep, GFP_KERNEL);
if (!pid)
return ERR_PTR(retval);
@@ -186,12 +189,31 @@ struct pid *alloc_pid(struct pid_namespace *ns)
if (idr_get_cursor(&tmp->idr) > RESERVED_PIDS)
pid_min = RESERVED_PIDS;

- /*
- * Store a null pointer so find_pid_ns does not find
- * a partially initialized PID (see below).
- */
- nr = idr_alloc_cyclic(&tmp->idr, NULL, pid_min,
- pid_max, GFP_ATOMIC);
+ if (set_tid) {
+ /*
+ * Also fail if a PID != 1 is requested
+ * and no PID 1 exists.
+ */
+ nr = -EINVAL;
+ if (set_tid == 1 || !idr_is_empty(&tmp->idr))
+ nr = idr_alloc(&tmp->idr, NULL, set_tid,
+ set_tid + 1, GFP_ATOMIC);
+ /*
+ * If ENOSPC is returned it means that the PID is
+ * alreay in use. Return EEXIST in that case.
+ */
+ if (nr == -ENOSPC)
+ nr = -EEXIST;
+ /* Only use set_tid for one PID namespace. */
+ set_tid = 0;
+ } else {
+ /*
+ * Store a null pointer so find_pid_ns does not find
+ * a partially initialized PID (see below).
+ */
+ nr = idr_alloc_cyclic(&tmp->idr, NULL, pid_min,
+ pid_max, GFP_ATOMIC);
+ }
spin_unlock_irq(&pidmap_lock);
idr_preload_end()

This makes things a lot more clearer in my opinion. First, verify that the
pre-conditions are met. Second, verify that the conditions are met which depend
on the state of the pid namespace, i.e. there's either already a pid 1 or pid 1
is requested.
We should also do this since alloc_pid() is exported in a header file and so we
can't and shouldn't rely on the fact that all callers will pass in something
sensible for set_tid.

>
> For consistency this should rather be:
>
> if (set_tid) {
> if ([...] && set_tid && [...])
> }
>
> Overall I'd rather write the second if as:
>
> if (set_tid && set_tid < pid_max && (set_tid == 1 || !idr_is_empty(&tmp->idr)))
>
> to place the most basic condition (i.e. set_tid is non-zero) leftmost.
>
> The right-most conjunct (set_tid == 1 || !idr_is_empty(&tmp->idr)) is
> non-trivial to parse. Someone not familiar with that code would need
> time to figure out why and how exactly 1 is special-cased.
> I think overall this is begging for alloc_pid() to be documented now if
> you don't mind. :)
>
> Christian