Re: [PATCH] coresight: etb10: restore atomic_t for shared reading state

[James Clark]

On 28/05/2026 5:52 pm, Runyu Xiao wrote:
> The etb10 miscdevice uses drvdata->reading as a shared exclusivity gate
> for userspace buffer access. etb_open() claims that gate with
> local_cmpxchg(), and etb_release() clears it with local_set().
>
> That gate is shared per-device state rather than CPU-local state. A
> running system can reach it whenever /dev/<etb> is opened, closed, and
> reopened by different tasks while the device remains registered, so the
> same drvdata->reading variable may be claimed on one CPU and later
> cleared on another.
>
> This code used to use atomic_t for the same gate, but commit
> 27b10da8fff2 ("coresight: etb10: moving to local atomic operations")
> changed it to local_t even though the access pattern remained cross-task
> and cross-CPU. Restore atomic_t together with atomic_cmpxchg() and
> atomic_set() so the exclusivity gate again uses a primitive intended
> for shared state.
>
> The issue was found on Linux v6.18.21 by our static analysis tool while
> scanning surviving local_t-on-shared-state sites, and then manually
> reviewed against the live etb10 file-op path.
>
> It was runtime-validated with a reproducible QEMU no-device KCSAN PoC
> that kept the same report-local contract:
>
>    1. use one shared struct etb_drvdata carrier and its
>       drvdata->reading gate;
>    2. call etb_open() and etb_release() sequentially on that gate to
>       confirm the original claim/clear path;
>    3. bind the open side to CPU0 and the release side to CPU1 for the
>       same gate to show cross-CPU ownership;
>    4. run bound workers that repeatedly race etb_open() and
>       etb_release() on the same gate until KCSAN reports a target hit.
>
> The harness recorded:
>
>    L1 passed open=1 release=1
>    reading_after_open=1 reading_after_release=0
>    L2 passed open_cpu=0 release_cpu=1
>    cross_cpu_release=1 reading_after=0 open_ret=0
>
> Representative KCSAN excerpt from the no-device validation run:
>
>    BUG: KCSAN: data-race in etb_open.constprop.0.isra.0 [vuln_msv]
>
>    write to 0xffffffffc0003810 of 4 bytes by task 216 on cpu 1:
>     etb_open.constprop.0.isra.0+0x38/0x80 [vuln_msv]
>     l3_worker_thread_fn+0x4f/0xf0 [vuln_msv]
>     kthread+0x17e/0x1c0
>     ret_from_fork+0x22/0x30
>
>    read to 0xffffffffc0003810 of 4 bytes by task 215 on cpu 0:
>     etb_open.constprop.0.isra.0+0x18/0x80 [vuln_msv]
>     l3_worker_thread_fn+0x4f/0xf0 [vuln_msv]
>     kthread+0x17e/0x1c0
>     ret_from_fork+0x22/0x30
>
>    value changed: 0x00000000 -> 0x00000001
>
>    Reported by Kernel Concurrency Sanitizer on:
>    CPU: 0 PID: 215 Comm: etb10_l3_a Tainted: G           O       6.1.66 #2
>
> This no-device harness is not a real ETB10 hardware end-to-end run, but
> it preserves the same shared drvdata->reading gate and the same
> etb_open()/etb_release() claim/clear contract. No real ETB10 hardware
> was available for runtime testing.
>
> Build-tested with:
>    make olddefconfig
>    make -j"$(nproc)" drivers/hwtracing/coresight/coresight-etb10.o
>
> Fixes: 27b10da8fff2 ("coresight: etb10: moving to local atomic operations")
> Cc: stable@xxxxxxxxxxxxxxx
> Signed-off-by: Runyu Xiao <runyu.xiao@xxxxxxxxxx>
> ---
>   drivers/hwtracing/coresight/coresight-etb10.c | 6 +++---
>   1 file changed, 3 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/hwtracing/coresight/coresight-etb10.c b/drivers/hwtracing/coresight/coresight-etb10.c
> index 35db1b6093d1..98269ea6f7ae 100644
> --- a/drivers/hwtracing/coresight/coresight-etb10.c
> +++ b/drivers/hwtracing/coresight/coresight-etb10.c
> @@ -85,7 +85,7 @@ struct etb_drvdata {
>   	struct coresight_device	*csdev;
>   	struct miscdevice	miscdev;
>   	raw_spinlock_t		spinlock;
> -	local_t			reading;
> +	atomic_t		reading;
>   	pid_t			pid;
>   	u8			*buf;
>   	u32			buffer_depth;
> @@ -603,7 +603,7 @@ static int etb_open(struct inode *inode, struct file *file)
>   	struct etb_drvdata *drvdata = container_of(file->private_data,
>   						   struct etb_drvdata, miscdev);
>   
> -	if (local_cmpxchg(&drvdata->reading, 0, 1))
> +	if (atomic_cmpxchg(&drvdata->reading, 0, 1))
>   		return -EBUSY;
>   
>   	dev_dbg(&drvdata->csdev->dev, "%s: successfully opened\n", __func__);
> @@ -641,7 +641,7 @@ static int etb_release(struct inode *inode, struct file *file)
>   {
>   	struct etb_drvdata *drvdata = container_of(file->private_data,
>   						   struct etb_drvdata, miscdev);
> -	local_set(&drvdata->reading, 0);
> +	atomic_set(&drvdata->reading, 0);
>   
>   	dev_dbg(&drvdata->csdev->dev, "%s: released\n", __func__);
>   	return 0;


Reviewed-by: James Clark <james.clark@xxxxxxxxxx>

Semi-related to this change, etb_read() doesn't have any lock when 
reading drvdata->buffer_dept or drvdata->buf. It locks in etb_dump(), 
but then unlocks before actually calling copy_to_user().

Seems like concurrent calls to etb_read() might end up with corrupt 
data, although I'm not sure if that would ever happen in practice 
because it only allows one open file handle.