[PATCH net v4] net/mlx5e: Fix a race in command alloc flow

[PATCH net v4] net/mlx5e: Fix a race in command alloc flow

[Shifeng Li]

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]
[70013.082028] 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
[70013.082173]  mlx5_cmd_trigger_completions+0x1e2/0x4c0 [mlx5_core]
[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]
[70013.082409]  pci_device_remove+0xf3/0x280
[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
[70013.082543]  process_one_work+0x65f/0x12d0
[70013.082556]  worker_thread+0x87/0xb50
[70013.082571]  kthread+0x2e9/0x3a0
[70013.082592]  ret_from_fork+0x1f/0x40

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().

Fixes: 50b2412b7e78 ("net/mlx5: Avoid possible free of command entry while timeout comp handler")
Reviewed-by: Moshe Shemesh <moshe@nvidia.com>
Signed-off-by: Shifeng Li <lishifeng@sangfor.com.cn>
---
 drivers/net/ethernet/mellanox/mlx5/core/cmd.c | 12 +++++++-----
 1 file changed, 7 insertions(+), 5 deletions(-)

---
v1->v2: fix code conflicts.
v2->v3: modify Fixes line and massage git log.
v3->v4: add target tree name in the subject and add the logical diagram.

diff --git a/drivers/net/ethernet/mellanox/mlx5/core/cmd.c b/drivers/net/ethernet/mellanox/mlx5/core/cmd.c
index f8f0a712c943..a7b1f9686c09 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/cmd.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/cmd.c
@@ -156,15 +156,18 @@ static u8 alloc_token(struct mlx5_cmd *cmd)
 	return token;
 }
 
-static int cmd_alloc_index(struct mlx5_cmd *cmd)
+static int cmd_alloc_index(struct mlx5_cmd *cmd, struct mlx5_cmd_work_ent *ent)
 {
 	unsigned long flags;
 	int ret;
 
 	spin_lock_irqsave(&cmd->alloc_lock, flags);
 	ret = find_first_bit(&cmd->vars.bitmask, cmd->vars.max_reg_cmds);
-	if (ret < cmd->vars.max_reg_cmds)
+	if (ret < cmd->vars.max_reg_cmds) {
 		clear_bit(ret, &cmd->vars.bitmask);
+		ent->idx = ret;
+		cmd->ent_arr[ent->idx] = ent;
+	}
 	spin_unlock_irqrestore(&cmd->alloc_lock, flags);
 
 	return ret < cmd->vars.max_reg_cmds ? ret : -ENOMEM;
@@ -979,7 +982,7 @@ static void cmd_work_handler(struct work_struct *work)
 	sem = ent->page_queue ? &cmd->vars.pages_sem : &cmd->vars.sem;
 	down(sem);
 	if (!ent->page_queue) {
-		alloc_ret = cmd_alloc_index(cmd);
+		alloc_ret = cmd_alloc_index(cmd, ent);
 		if (alloc_ret < 0) {
 			mlx5_core_err_rl(dev, "failed to allocate command entry\n");
 			if (ent->callback) {
@@ -994,15 +997,14 @@ static void cmd_work_handler(struct work_struct *work)
 			up(sem);
 			return;
 		}
-		ent->idx = alloc_ret;
 	} else {
 		ent->idx = cmd->vars.max_reg_cmds;
 		spin_lock_irqsave(&cmd->alloc_lock, flags);
 		clear_bit(ent->idx, &cmd->vars.bitmask);
+		cmd->ent_arr[ent->idx] = ent;
 		spin_unlock_irqrestore(&cmd->alloc_lock, flags);
 	}
 
-	cmd->ent_arr[ent->idx] = ent;
 	lay = get_inst(cmd, ent->idx);
 	ent->lay = lay;
 	memset(lay, 0, sizeof(*lay));
-- 
2.25.1

Re: [PATCH net v4] net/mlx5e: Fix a race in command alloc flow

[Saeed Mahameed]

On 02 Dec 00:01, 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]
>[70013.082028] 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
>[70013.082173]  mlx5_cmd_trigger_completions+0x1e2/0x4c0 [mlx5_core]
>[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]
>[70013.082409]  pci_device_remove+0xf3/0x280
>[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
>[70013.082543]  process_one_work+0x65f/0x12d0
>[70013.082556]  worker_thread+0x87/0xb50
>[70013.082571]  kthread+0x2e9/0x3a0
>[70013.082592]  ret_from_fork+0x1f/0x40
>
>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().
>
>Fixes: 50b2412b7e78 ("net/mlx5: Avoid possible free of command entry while timeout comp handler")
>Reviewed-by: Moshe Shemesh <moshe@nvidia.com>
>Signed-off-by: Shifeng Li <lishifeng@sangfor.com.cn>


LGTM,
Applied to net-mlx5.

Thanks,
Saeed.