[Devel] Re: [PATCH 5/6] c/r: correctly restore pgid

[Serge E. Hallyn]

Quoting Oren Laadan (orenl at librato.com):
> 
> 
> Serge E. Hallyn wrote:
> > Quoting Oren Laadan (orenl at librato.com):
> >> The main challenge with restoring the pgid of tasks is that the
> >> original "owner" (the process with that pid) might have exited
> >> already. I call these "ghost" pgids. 'mktree' does create these
> >> processes, but they then exit without participating in the restart.
> >>
> >> To solve this, this patch introduces a RESTART_GHOST flag, used for
> >> "ghost" owners that are created only to pass their pgid to other
> >> tasks. ('mktree' now makes them call restart(2) instead of exiting).
> >>
> >> When a "ghost" task calls restart(2), it will be placed on a wait
> >> queue until the restart completes and then exit. This guarantees that
> >> the pgid that it owns remains available for all (regular) restarting
> >> tasks for when they need it.
> >>
> >> Regular tasks perform the restart as before, except that they also
> >> now restore their old pgrp, which is guaranteed to exist.
> >>
> >> Changelog [v1]:
> >>   - Verify that pgid owner is a thread-group-leader.
> >>   - Handle the case of pgid/sid == 0 using root's parent pid-ns
> >>
> >> Signed-off-by: Oren Laadan <orenl at cs.colubmia.edu>
> >> ---
> >>  checkpoint/process.c             |  106 ++++++++++++++++++++++++-
> >>  checkpoint/restart.c             |  158 ++++++++++++++++++++++++++------------
> >>  checkpoint/sys.c                 |    3 +-
> >>  include/linux/checkpoint.h       |   11 ++-
> >>  include/linux/checkpoint_hdr.h   |    3 +
> >>  include/linux/checkpoint_types.h |    6 +-
> >>  6 files changed, 230 insertions(+), 57 deletions(-)
> >>
> >> diff --git a/checkpoint/process.c b/checkpoint/process.c
> >> index 40b2580..5d6bdb9 100644
> >> --- a/checkpoint/process.c
> >> +++ b/checkpoint/process.c
> >> @@ -23,6 +23,57 @@
> >>  #include <linux/syscalls.h>
> >>
> >>
> >> +pid_t ckpt_pid_nr(struct ckpt_ctx *ctx, struct pid *pid)
> >> +{
> >> +	return pid ? pid_nr_ns(pid, ctx->root_nsproxy->pid_ns) : CKPT_PID_NULL;
> >> +}
> >> +
> >> +/* must be called with tasklist_lock or rcu_read_lock() held */
> >> +struct pid *_ckpt_find_pgrp(struct ckpt_ctx *ctx, pid_t pgid)
> >> +{
> >> +	struct task_struct *p;
> >> +	struct pid *pgrp;
> >> +
> >> +	if (pgid == 0) {
> >> +		/*
> >> +		 * At checkpoint the pgid owner lived in an ancestor
> >> +		 * pid-ns. The best we can do (sanely and safely) is
> >> +		 * to examine the parent of this restart's root: if in
> >> +		 * a distinct pid-ns, use its pgrp; otherwise fail.
> >> +		 */
> >> +		p = ctx->root_task->real_parent;
> >> +		if (p->nsproxy->pid_ns == current->nsproxy->pid_ns)
> >> +			return NULL;
> >> +		pgrp = task_pgrp(p);
> >> +	} else {
> >> +		/*
> >> +		 * Find the owner process of this pgid (it must exist
> >> +		 * if pgrp exists). It must be a thread group leader.
> >> +		 */
> >> +		pgrp = find_vpid(pgid);
> >> +		p = pid_task(pgrp, PIDTYPE_PID);
> >> +		if (!p || !thread_group_leader(p))
> >> +			return NULL;
> >> +		/*
> >> +		 * The pgrp must "belong" to our restart tree (compare
> >> +		 * p->checkpoint_ctx to ours). This prevents malicious
> >> +		 * input from (guessing and) using unrelated pgrps. If
> >> +		 * the owner is dead, then it doesn't have a context,
> >> +		 * so instead compare against its (real) parent's.
> >> +		 */
> >> +		if (p->exit_state == EXIT_ZOMBIE)
> >> +			p = p->real_parent;
> >> +		if (p->checkpoint_ctx != ctx)
> >> +			return NULL;
> >> +	}
> >> +
> >> +	if (task_session(current) != task_session(p))
> >> +		return NULL;
> >> +
> >> +	return pgrp;
> >> +}
> >> +
> >> +
> >>  #ifdef CONFIG_FUTEX
> >>  static void save_task_robust_futex_list(struct ckpt_hdr_task *h,
> >>  					struct task_struct *t)
> >> @@ -94,8 +145,8 @@ static int checkpoint_task_struct(struct ckpt_ctx *ctx, struct task_struct *t)
> >>  		h->exit_signal = t->exit_signal;
> >>  		h->pdeath_signal = t->pdeath_signal;
> >>
> >> -		h->set_child_tid = t->set_child_tid;
> >> -		h->clear_child_tid = t->clear_child_tid;
> >> +		h->set_child_tid = (unsigned long) t->set_child_tid;
> > 
> > note that set_child_tid is an int (signed), not a long.  Same on
> > x86, but not on other arches.  Shouldn't lose info so could be worse.
> 
> {set,clear}_child_tid are both pointers to user space: it's an address
> in userspace, so we save it as 'unsigned long'.
> 
> {clear,set}_child_tid is defined in include/linux/sched.h ... how can
> it differ for different archs ?

sizeof long differs for different archs.  Not the type of x_child_tid.

> > 
> > On the whole,
> > 
> > Acked-by: Serge Hallyn <serue at us.ibm.com>
> 
> Thanks. I got a few fixes for the code piles up and now c/r of 'screen'
> with a couple of shells is working :)

Cool!
_______________________________________________
Containers mailing list
Containers at lists.linux-foundation.org
https://lists.linux-foundation.org/mailman/listinfo/containers