Re: [PATCH v2] net/mlx5e: Fix a race in command alloc flow

[Shifeng Li]

On 2023/11/22 20:02, Leon Romanovsky wrote:
> On Tue, Nov 21, 2023 at 03:52:51AM -0800, Shifeng Li wrote:
> > Fix a cmd->ent use after free due to a race on command entry.
> > Such race occurs when one of the commands releases its last refcount and
> > frees its index and entry while another process running command flush
> > flow takes refcount to this command entry. The process which handles
> > commands flush may see this command as needed to be flushed if the other
> > process allocated a ent->idx but didn't set ent to cmd->ent_arr in
> > cmd_work_handler(). Fix it by moving the assignment of cmd->ent_arr into
> > the spin lock.
> >
> > [70013.081955] BUG: KASAN: use-after-free in mlx5_cmd_trigger_completions+0x1e2/0x4c0 [mlx5_core]
> > [70013.081967] Write of size 4 at addr ffff88880b1510b4 by task kworker/26:1/1433361
> > [70013.081968]
> > [70013.081989] CPU: 26 PID: 1433361 Comm: kworker/26:1 Kdump: loaded Tainted: G           OE     4.19.90-25.17.v2101.osc.sfc.6.10.0.0030.ky10.x86_64+debug #1
> > [70013.082001] Hardware name: SANGFOR 65N32-US/ASERVER-G-2605, BIOS SSSS5203 08/19/2020
> > [70013.082028] Workqueue: events aer_isr
> > [70013.082053] Call Trace:
> > [70013.082067]  dump_stack+0x8b/0xbb
> > [70013.082086]  print_address_description+0x6a/0x270
> > [70013.082102]  kasan_report+0x179/0x2c0
> > [70013.082133]  ? mlx5_cmd_trigger_completions+0x1e2/0x4c0 [mlx5_core]
> > [70013.082173]  mlx5_cmd_trigger_completions+0x1e2/0x4c0 [mlx5_core]
> > [70013.082213]  ? mlx5_cmd_use_polling+0x20/0x20 [mlx5_core]
> > [70013.082223]  ? kmem_cache_free+0x1ad/0x1e0
> > [70013.082267]  mlx5_cmd_flush+0x80/0x180 [mlx5_core]
> > [70013.082304]  mlx5_enter_error_state+0x106/0x1d0 [mlx5_core]
> > [70013.082338]  mlx5_try_fast_unload+0x2ea/0x4d0 [mlx5_core]
> > [70013.082377]  remove_one+0x200/0x2b0 [mlx5_core]
> > [70013.082390]  ? __pm_runtime_resume+0x58/0x70
> > [70013.082409]  pci_device_remove+0xf3/0x280
> > [70013.082426]  ? pcibios_free_irq+0x10/0x10
> > [70013.082439]  device_release_driver_internal+0x1c3/0x470
> > [70013.082453]  pci_stop_bus_device+0x109/0x160
> > [70013.082468]  pci_stop_and_remove_bus_device+0xe/0x20
> > [70013.082485]  pcie_do_fatal_recovery+0x167/0x550
> > [70013.082493]  aer_isr+0x7d2/0x960
> > [70013.082510]  ? aer_get_device_error_info+0x420/0x420
> > [70013.082526]  ? __schedule+0x821/0x2040
> > [70013.082536]  ? strscpy+0x85/0x180
> > [70013.082543]  process_one_work+0x65f/0x12d0
> > [70013.082556]  worker_thread+0x87/0xb50
> > [70013.082563]  ? __kthread_parkme+0x82/0xf0
> > [70013.082569]  ? process_one_work+0x12d0/0x12d0
> > [70013.082571]  kthread+0x2e9/0x3a0
> > [70013.082579]  ? kthread_create_worker_on_cpu+0xc0/0xc0
> > [70013.082592]  ret_from_fork+0x1f/0x40
>
> I'm curious how did you get this error? I would expect to see some sort
> of lock in upper level which prevents it.
The logical relationship of this error is as follows:

                  aer_recover_work                    |             ent->work
------------------------------------------------------+---------------------------------
aer_recover_work_func                                 |
|- pcie_do_recovery                                   |
  |- report_error_detected                            |
    |- mlx5_pci_err_detected                          |cmd_work_handler
      |- mlx5_enter_error_state                       |  |- cmd_alloc_index
        |- enter_error_state                          |    |- lock cmd->alloc_lock
          |- mlx5_cmd_flush                           |    |- clear_bit
            |- mlx5_cmd_trigger_completions           |    |- unlock cmd->alloc_lock
              |- lock cmd->alloc_lock                 |
              |- vector = ~dev->cmd.vars.bitmask      |
              |- for_each_set_bit                     |
                |- cmd_ent_get(cmd->ent_arr[i]) (UAF) |
              |- unlock cmd->alloc_lock               |  |- cmd->ent_arr[ent->idx] = ent

The cmd->ent_arr[ent->idx] assignment and the bit clearing are not protected by the cmd->alloc_lock in cmd_work_handler().

> Thanks