devlink: Break parameter notification sequence to be before/after unload/load driver
authorLeon Romanovsky <leonro@nvidia.com>
Thu, 29 Jul 2021 17:19:24 +0000 (20:19 +0300)
committerJakub Kicinski <kuba@kernel.org>
Fri, 30 Jul 2021 20:16:38 +0000 (13:16 -0700)
The change of namespaces during devlink reload calls to driver unload
before it accesses devlink parameters. The commands below causes to
use-after-free bug when trying to get flow steering mode.

 * ip netns add n1
 * devlink dev reload pci/0000:00:09.0 netns n1

 ==================================================================
 BUG: KASAN: use-after-free in mlx5_devlink_fs_mode_get+0x96/0xa0 [mlx5_core]
 Read of size 4 at addr ffff888009d04308 by task devlink/275

 CPU: 6 PID: 275 Comm: devlink Not tainted 5.12.0-rc2+ #2853
 Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.13.0-0-gf21b5a4aeb02-prebuilt.qemu.org 04/01/2014
 Call Trace:
  dump_stack+0x93/0xc2
  print_address_description.constprop.0+0x18/0x140
  ? mlx5_devlink_fs_mode_get+0x96/0xa0 [mlx5_core]
  ? mlx5_devlink_fs_mode_get+0x96/0xa0 [mlx5_core]
  kasan_report.cold+0x7c/0xd8
  ? mlx5_devlink_fs_mode_get+0x96/0xa0 [mlx5_core]
  mlx5_devlink_fs_mode_get+0x96/0xa0 [mlx5_core]
  devlink_nl_param_fill+0x1c8/0xe80
  ? __free_pages_ok+0x37a/0x8a0
  ? devlink_flash_update_timeout_notify+0xd0/0xd0
  ? lock_acquire+0x1a9/0x6d0
  ? fs_reclaim_acquire+0xb7/0x160
  ? lock_is_held_type+0x98/0x110
  ? 0xffffffff81000000
  ? lock_release+0x1f9/0x6c0
  ? fs_reclaim_release+0xa1/0xf0
  ? lock_downgrade+0x6d0/0x6d0
  ? lock_is_held_type+0x98/0x110
  ? lock_is_held_type+0x98/0x110
  ? memset+0x20/0x40
  ? __build_skb_around+0x1f8/0x2b0
  devlink_param_notify+0x6d/0x180
  devlink_reload+0x1c3/0x520
  ? devlink_remote_reload_actions_performed+0x30/0x30
  ? mutex_trylock+0x24b/0x2d0
  ? devlink_nl_cmd_reload+0x62b/0x1070
  devlink_nl_cmd_reload+0x66d/0x1070
  ? devlink_reload+0x520/0x520
  ? devlink_get_from_attrs+0x1bc/0x260
  ? devlink_nl_pre_doit+0x64/0x4d0
  genl_family_rcv_msg_doit+0x1e9/0x2f0
  ? mutex_lock_io_nested+0x1130/0x1130
  ? genl_family_rcv_msg_attrs_parse.constprop.0+0x240/0x240
  ? security_capable+0x51/0x90
  genl_rcv_msg+0x27f/0x4a0
  ? genl_get_cmd+0x3c0/0x3c0
  ? lock_acquire+0x1a9/0x6d0
  ? devlink_reload+0x520/0x520
  ? lock_release+0x6c0/0x6c0
  netlink_rcv_skb+0x11d/0x340
  ? genl_get_cmd+0x3c0/0x3c0
  ? netlink_ack+0x9f0/0x9f0
  ? lock_release+0x1f9/0x6c0
  genl_rcv+0x24/0x40
  netlink_unicast+0x433/0x700
  ? netlink_attachskb+0x730/0x730
  ? _copy_from_iter_full+0x178/0x650
  ? __alloc_skb+0x113/0x2b0
  netlink_sendmsg+0x6f1/0xbd0
  ? netlink_unicast+0x700/0x700
  ? lock_is_held_type+0x98/0x110
  ? netlink_unicast+0x700/0x700
  sock_sendmsg+0xb0/0xe0
  __sys_sendto+0x193/0x240
  ? __x64_sys_getpeername+0xb0/0xb0
  ? do_sys_openat2+0x10b/0x370
  ? __up_read+0x1a1/0x7b0
  ? do_user_addr_fault+0x219/0xdc0
  ? __x64_sys_openat+0x120/0x1d0
  ? __x64_sys_open+0x1a0/0x1a0
  __x64_sys_sendto+0xdd/0x1b0
  ? syscall_enter_from_user_mode+0x1d/0x50
  do_syscall_64+0x2d/0x40
  entry_SYSCALL_64_after_hwframe+0x44/0xae
 RIP: 0033:0x7fc69d0af14a
 Code: d8 64 89 02 48 c7 c0 ff ff ff ff eb b8 0f 1f 00 f3 0f 1e fa 41 89 ca 64 8b 04 25 18 00 00 00 85 c0 75 15 b8 2c 00 00 00 0f 05 <48> 3d 00 f0 ff ff 77 76 c3 0f 1f 44 00 00 55 48 83 ec 30 44 89 4c
 RSP: 002b:00007ffc1d8292f8 EFLAGS: 00000246 ORIG_RAX: 000000000000002c
 RAX: ffffffffffffffda RBX: 0000000000000005 RCX: 00007fc69d0af14a
 RDX: 0000000000000038 RSI: 0000555f57c56440 RDI: 0000000000000003
 RBP: 0000555f57c56410 R08: 00007fc69d17b200 R09: 000000000000000c
 R10: 0000000000000000 R11: 0000000000000246 R12: 0000000000000000
 R13: 0000000000000000 R14: 0000000000000000 R15: 0000000000000000

 Allocated by task 146:
  kasan_save_stack+0x1b/0x40
  __kasan_kmalloc+0x99/0xc0
  mlx5_init_fs+0xf0/0x1c50 [mlx5_core]
  mlx5_load+0xd2/0x180 [mlx5_core]
  mlx5_init_one+0x2f6/0x450 [mlx5_core]
  probe_one+0x47d/0x6e0 [mlx5_core]
  pci_device_probe+0x2a0/0x4a0
  really_probe+0x20a/0xc90
  driver_probe_device+0xd8/0x380
  device_driver_attach+0x1df/0x250
  __driver_attach+0xff/0x240
  bus_for_each_dev+0x11e/0x1a0
  bus_add_driver+0x309/0x570
  driver_register+0x1ee/0x380
  0xffffffffa06b8062
  do_one_initcall+0xd5/0x410
  do_init_module+0x1c8/0x760
  load_module+0x6d8b/0x9650
  __do_sys_finit_module+0x118/0x1b0
  do_syscall_64+0x2d/0x40
  entry_SYSCALL_64_after_hwframe+0x44/0xae

 Freed by task 275:
  kasan_save_stack+0x1b/0x40
  kasan_set_track+0x1c/0x30
  kasan_set_free_info+0x20/0x30
  __kasan_slab_free+0x102/0x140
  slab_free_freelist_hook+0x74/0x1b0
  kfree+0xd7/0x2a0
  mlx5_unload+0x16/0xb0 [mlx5_core]
  mlx5_unload_one+0xae/0x120 [mlx5_core]
  mlx5_devlink_reload_down+0x1bc/0x380 [mlx5_core]
  devlink_reload+0x141/0x520
  devlink_nl_cmd_reload+0x66d/0x1070
  genl_family_rcv_msg_doit+0x1e9/0x2f0
  genl_rcv_msg+0x27f/0x4a0
  netlink_rcv_skb+0x11d/0x340
  genl_rcv+0x24/0x40
  netlink_unicast+0x433/0x700
  netlink_sendmsg+0x6f1/0xbd0
  sock_sendmsg+0xb0/0xe0
  __sys_sendto+0x193/0x240
  __x64_sys_sendto+0xdd/0x1b0
  do_syscall_64+0x2d/0x40
  entry_SYSCALL_64_after_hwframe+0x44/0xae

 The buggy address belongs to the object at ffff888009d04300
  which belongs to the cache kmalloc-128 of size 128
 The buggy address is located 8 bytes inside of
  128-byte region [ffff888009d04300ffff888009d04380)
 The buggy address belongs to the page:
 page:0000000086a64ecc refcount:1 mapcount:0 mapping:0000000000000000 index:0xffff888009d04000 pfn:0x9d04
 head:0000000086a64ecc order:1 compound_mapcount:0
 flags: 0x4000000000010200(slab|head)
 raw: 4000000000010200 ffffea0000203980 0000000200000002 ffff8880050428c0
 raw: ffff888009d04000 000000008020001d 00000001ffffffff 0000000000000000
 page dumped because: kasan: bad access detected

 Memory state around the buggy address:
  ffff888009d04200: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb
  ffff888009d04280: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc
 >ffff888009d04300: fa fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb
                       ^
  ffff888009d04380: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc
  ffff888009d04400: fa fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb
 ==================================================================

The right solution to devlink reload is to notify about deletion of
parameters, unload driver, change net namespaces, load driver and notify
about addition of parameters.

Fixes: 070c63f20f6c ("net: devlink: allow to change namespaces during reload")
Reviewed-by: Parav Pandit <parav@nvidia.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
net/core/devlink.c

index b596a97..fbca61a 100644 (file)
@@ -3801,10 +3801,12 @@ static void devlink_param_notify(struct devlink *devlink,
                                 struct devlink_param_item *param_item,
                                 enum devlink_command cmd);
 
-static void devlink_reload_netns_change(struct devlink *devlink,
-                                       struct net *dest_net)
+static void devlink_ns_change_notify(struct devlink *devlink,
+                                    struct net *dest_net, struct net *curr_net,
+                                    bool new)
 {
        struct devlink_param_item *param_item;
+       enum devlink_command cmd;
 
        /* Userspace needs to be notified about devlink objects
         * removed from original and entering new network namespace.
@@ -3812,17 +3814,18 @@ static void devlink_reload_netns_change(struct devlink *devlink,
         * reload process so the notifications are generated separatelly.
         */
 
-       list_for_each_entry(param_item, &devlink->param_list, list)
-               devlink_param_notify(devlink, 0, param_item,
-                                    DEVLINK_CMD_PARAM_DEL);
-       devlink_notify(devlink, DEVLINK_CMD_DEL);
+       if (!dest_net || net_eq(dest_net, curr_net))
+               return;
 
-       __devlink_net_set(devlink, dest_net);
+       if (new)
+               devlink_notify(devlink, DEVLINK_CMD_NEW);
 
-       devlink_notify(devlink, DEVLINK_CMD_NEW);
+       cmd = new ? DEVLINK_CMD_PARAM_NEW : DEVLINK_CMD_PARAM_DEL;
        list_for_each_entry(param_item, &devlink->param_list, list)
-               devlink_param_notify(devlink, 0, param_item,
-                                    DEVLINK_CMD_PARAM_NEW);
+               devlink_param_notify(devlink, 0, param_item, cmd);
+
+       if (!new)
+               devlink_notify(devlink, DEVLINK_CMD_DEL);
 }
 
 static bool devlink_reload_supported(const struct devlink_ops *ops)
@@ -3902,6 +3905,7 @@ static int devlink_reload(struct devlink *devlink, struct net *dest_net,
                          u32 *actions_performed, struct netlink_ext_ack *extack)
 {
        u32 remote_reload_stats[DEVLINK_RELOAD_STATS_ARRAY_SIZE];
+       struct net *curr_net;
        int err;
 
        if (!devlink->reload_enabled)
@@ -3909,18 +3913,22 @@ static int devlink_reload(struct devlink *devlink, struct net *dest_net,
 
        memcpy(remote_reload_stats, devlink->stats.remote_reload_stats,
               sizeof(remote_reload_stats));
+
+       curr_net = devlink_net(devlink);
+       devlink_ns_change_notify(devlink, dest_net, curr_net, false);
        err = devlink->ops->reload_down(devlink, !!dest_net, action, limit, extack);
        if (err)
                return err;
 
-       if (dest_net && !net_eq(dest_net, devlink_net(devlink)))
-               devlink_reload_netns_change(devlink, dest_net);
+       if (dest_net && !net_eq(dest_net, curr_net))
+               __devlink_net_set(devlink, dest_net);
 
        err = devlink->ops->reload_up(devlink, action, limit, actions_performed, extack);
        devlink_reload_failed_set(devlink, !!err);
        if (err)
                return err;
 
+       devlink_ns_change_notify(devlink, dest_net, curr_net, true);
        WARN_ON(!(*actions_performed & BIT(action)));
        /* Catch driver on updating the remote action within devlink reload */
        WARN_ON(memcmp(remote_reload_stats, devlink->stats.remote_reload_stats,