Use of rte_memseg_contig_walk in vhost_kernel.c causes deadlock

https://bugs.dpdk.org/show_bug.cgi?id=81

            Bug ID: 81
           Summary: Use of rte_memseg_contig_walk in vhost_kernel.c causes
                    deadlock
           Product: DPDK
           Version: 18.08
          Hardware: All
                OS: All
            Status: CONFIRMED
          Severity: normal
          Priority: Normal
         Component: vhost/virtio
          Assignee: [email protected]
          Reporter: [email protected]

Target Milestone: ---

Deadlock can occur when allocating memory if a virtio_user device is in use. The execution path that causes is this is as follows:

alloc_more_mem_on_socket(...);
try_expand_heap(...);

    // RW Write Lock taken here...
    rte_rwlock_write_lock(&mcfg->memory_hotplug_lock);

    try_expand_heap_primary(...);
      eal_memalloc_mem_event_notify(...);
        virtio_user_mem_event_cb(...);
          vhost_kernel_ioctl(...);
            prepare_vhost_memory_kernel(...);
              rte_memseg_contig_walk(...);

                // Attempt to take RW Read Lock here...
                rte_rwlock_read_lock(&mcfg->memory_hotplug_lock);

Replacing rte_memseg_contig_walk() with rte_memseg_walk_thread_unsafe() in prepare_vhost_memory_kernel() will remove the deadlock, but could cause issues as prepare_vhost_memory_kernel() may also be called outside of memalloc event notify context.

You are receiving this mail because:
You are the assignee for the bug.