Skip to contentSkip to conversation

mlx5: pthread mutex in shared memory missing PTHREAD_PROCESS_SHARED

DPDK/ethdev Bug 1875 [email protected] 1 participant 1 message Jan 30

lore
Share & tools
lore
#0 lore raw

http://bugs.dpdk.org/show_bug.cgi?id=1875

            Bug ID: 1875
           Summary: mlx5: pthread mutex in shared memory missing
                    PTHREAD_PROCESS_SHARED
           Product: DPDK
           Version: 25.11
          Hardware: All
                OS: All
            Status: UNCONFIRMED
          Severity: normal
          Priority: Normal
         Component: ethdev
          Assignee: [email protected]
          Reporter: [email protected]

Target Milestone: ---

This is the mlx5 subset of bugzilla 662

The mlx5 specific analysis was:

## MLX5 Driver PTHREAD_PROCESS_SHARED Issue

### Summary

The mlx5 driver has pthread mutexes in shared memory that are initialized without PTHREAD_PROCESS_SHARED, which causes undefined behavior when used across multiple processes.

### Affected Mutexes

1. **sh->txpp.mutex** (drivers/net/mlx5/mlx5.c, line 1846)
- Location: struct mlx5_dev_txpp inside struct mlx5_dev_ctx_shared
- Purpose: Pacing create/destroy mutex for Tx packet pacing

2. **sh->lwm_config_lock** (drivers/net/mlx5/mlx5.c, line 1963)
- Location: struct mlx5_dev_ctx_shared
- Purpose: Low watermark configuration lock

### Why This Is a Problem

The mlx5_dev_ctx_shared structure is allocated via mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE, ...) which uses hugepage-backed shared memory. This shared context is accessible by both primary and secondary processes.

While the shared context is only **allocated** in the primary process (enforced by assertion at line 1826), secondary processes can obtain a reference to this shared context and potentially call operations that lock these mutexes.

Per POSIX, mutexes in shared memory accessed by multiple processes must be initialized with PTHREAD_PROCESS_SHARED attribute. Without this, synchronization between processes is undefined behavior.

### Note

A third mutex, mlx5_dev_ctx_list_mutex (line 201), is a static global variable and is **not affected** since each process has its own copy.

### Suggested Fix

Initialize the mutexes with PTHREAD_PROCESS_SHARED:

```c
static void
mlx5_init_shared_mutex(pthread_mutex_t *mutex)
{
    pthread_mutexattr_t attr;

    pthread_mutexattr_init(&attr);
    pthread_mutexattr_setpshared(&attr, PTHREAD_PROCESS_SHARED);
    pthread_mutex_init(mutex, &attr);
    pthread_mutexattr_destroy(&attr);
}
```

Then replace:

```c
pthread_mutex_init(&sh->txpp.mutex, NULL);
pthread_mutex_init(&priv->sh->lwm_config_lock, NULL);
```

With:

```c
mlx5_init_shared_mutex(&sh->txpp.mutex);
mlx5_init_shared_mutex(&priv->sh->lwm_config_lock);
```
Signature · 2 lines
You are receiving this mail because:
You are the assignee for the bug.