Re: [PATCH] mm/mempolicy.c: Remove unnecessary nodemask check in kernel_migrate_pages()

On 2019/8/6 16:36, Vlastimil Babka wrote:

On 8/6/19 4:36 AM, Kefeng Wang wrote:

[...]

quoted

[QUESTION]

SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
                const unsigned long __user *, old_nodes,
                const unsigned long __user *, new_nodes)
{
        return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes);
}

The migrate_pages() takes pid argument, witch is the ID of the process
whose pages are to be moved. should the cpuset_mems_allowed(current) be
cpuset_mems_allowed(task)?

The check for cpuset_mems_allowed(task) is just above the code you change, so
the new nodes have to be subset of the target task's cpuset.
But they also have to be allowed by the calling task's cpuset. In manpage of
migrate_pages(2), this is hinted by the NOTES "Use get_mempolicy(2) with the
MPOL_F_MEMS_ALLOWED flag to obtain the set of nodes that are allowed by the
calling process's cpuset..."

But perhaps the manpage should be better clarified:

- the EINVAL case includes "Or, none of the node IDs specified by new_nodes are
on-line and allowed by the process's current cpuset context, or none of the
specified nodes contain memory." - this should probably say "calling process" to
disambiguate
- the EPERM case should mention that new_nodes have to be subset of the target
process' cpuset context. The caller should also have CAP_SYS_NICE and
ptrace_may_access()

Get it, thanks for your detail explanation.