Re: [PATCH linux dev-4.13 10/16] hwmon (p9_sbe): Convert client_lock from a spinlock to a mutex

[Eddie James <eajames@linux.vnet.ibm.com>]

On 02/19/2018 10:18 PM, Andrew Jeffery wrote:
> The hwmon occ implementation allocates GFP_KERNEL memory in occ_open_common(),
> therefore we cannot call it under a spinlock as it may sleep. Compiling with
> lock debugging enabled gives us the following warning:

Acked-by: Eddie James <eajames@linux.vnet.ibm.com>

>
> [    4.420000] WARNING: CPU: 0 PID: 1 at kernel/locking/lockdep.c:2879 lockdep_trace_alloc+0xf0/0x124
> [    4.420000] DEBUG_LOCKS_WARN_ON(irqs_disabled_flags(flags))
> [    4.420000] CPU: 0 PID: 1 Comm: swapper Not tainted 4.10.17-00526-g1cbacc6bd3a1 #2334
> [    4.420000] Hardware name: ASpeed SoC
> [    4.420000] [<80010eec>] (unwind_backtrace) from [<8000e664>] (show_stack+0x20/0x24)
> [    4.420000] [<8000e664>] (show_stack) from [<80249160>] (dump_stack+0x20/0x28)
> [    4.420000] [<80249160>] (dump_stack) from [<8001d3a0>] (__warn+0xe0/0x108)
> [    4.420000] [<8001d3a0>] (__warn) from [<8001d40c>] (warn_slowpath_fmt+0x44/0x4c)
> [    4.420000] [<8001d40c>] (warn_slowpath_fmt) from [<8005d2cc>] (lockdep_trace_alloc+0xf0/0x124)
> [    4.420000] [<8005d2cc>] (lockdep_trace_alloc) from [<8012bb1c>] (kmem_cache_alloc_trace+0x3c/0x284)
> [    4.420000] [<8012bb1c>] (kmem_cache_alloc_trace) from [<8034fa2c>] (occ_open_common+0x30/0xac)
> [    4.420000] [<8034fa2c>] (occ_open_common) from [<80350bac>] (occ_drv_open+0x24/0x28)
> [    4.420000] [<80350bac>] (occ_drv_open) from [<803638fc>] (p9_sbe_occ_send_cmd+0x44/0x13c)
> [    4.420000] [<803638fc>] (p9_sbe_occ_send_cmd) from [<803615b8>] (occ_poll+0x6c/0x1c0)
> [    4.420000] [<803615b8>] (occ_poll) from [<80363a84>] (p9_sbe_occ_probe+0x90/0x178)
> [    4.420000] [<80363a84>] (p9_sbe_occ_probe) from [<802b9850>] (platform_drv_probe+0x60/0xbc)
> [    4.420000] [<802b9850>] (platform_drv_probe) from [<802b76b0>] (driver_probe_device+0x114/0x430)
> [    4.420000] [<802b76b0>] (driver_probe_device) from [<802b7a98>] (__driver_attach+0xcc/0x10c)
> [    4.420000] [<802b7a98>] (__driver_attach) from [<802b5770>] (bus_for_each_dev+0x5c/0xac)
> [    4.420000] [<802b5770>] (bus_for_each_dev) from [<802b7c74>] (driver_attach+0x28/0x30)
> [    4.420000] [<802b7c74>] (driver_attach) from [<802b6268>] (bus_add_driver+0x19c/0x25c)
> [    4.420000] [<802b6268>] (bus_add_driver) from [<802b8abc>] (driver_register+0x88/0x104)
> [    4.420000] [<802b8abc>] (driver_register) from [<802ba468>] (__platform_driver_register+0x3c/0x50)
> [    4.420000] [<802ba468>] (__platform_driver_register) from [<8066d804>] (p9_sbe_occ_driver_init+0x18/0x20)
> [    4.420000] [<8066d804>] (p9_sbe_occ_driver_init) from [<8064ae7c>] (do_one_initcall+0xa8/0x168)
> [    4.420000] [<8064ae7c>] (do_one_initcall) from [<8064b04c>] (kernel_init_freeable+0x110/0x1c8)
> [    4.420000] [<8064b04c>] (kernel_init_freeable) from [<804a9034>] (kernel_init+0x18/0x104)
> [    4.420000] [<804a9034>] (kernel_init) from [<8000a888>] (ret_from_fork+0x14/0x2c)
>
> Avoid the warning (or the need for GFP_ATOMIC) and allow for reduced system
> latency by using a mutex instead. Mutual exclusion is the only requirement, we
> do not need to block pre-emption.
>
> Signed-off-by: Andrew Jeffery <andrew@aj.id.au>
> ---
>   drivers/hwmon/occ/p9_sbe.c | 14 ++++++--------
>   1 file changed, 6 insertions(+), 8 deletions(-)
>
> diff --git a/drivers/hwmon/occ/p9_sbe.c b/drivers/hwmon/occ/p9_sbe.c
> index 51bdfa89f1a6..38cf57ad1bb8 100644
> --- a/drivers/hwmon/occ/p9_sbe.c
> +++ b/drivers/hwmon/occ/p9_sbe.c
> @@ -34,34 +34,32 @@ struct p9_sbe_occ {
>   	 * open, close and NULL assignment. This prevents simultaneous opening
>   	 * and closing of the client, or closing multiple times.
>   	 */
> -	spinlock_t client_lock;
> +	struct mutex client_lock;
>   };
>
>   #define to_p9_sbe_occ(x)	container_of((x), struct p9_sbe_occ, occ)
>
>   static void p9_sbe_occ_close_client(struct p9_sbe_occ *ctx)
>   {
> -	unsigned long flags;
>   	struct occ_client *tmp_client;
>
> -	spin_lock_irqsave(&ctx->client_lock, flags);
> +	mutex_lock(&ctx->client_lock);
>   	tmp_client = ctx->client;
>   	ctx->client = NULL;
>   	occ_drv_release(tmp_client);
> -	spin_unlock_irqrestore(&ctx->client_lock, flags);
> +	mutex_unlock(&ctx->client_lock);
>   }
>
>   static int p9_sbe_occ_send_cmd(struct occ *occ, u8 *cmd)
>   {
>   	int rc;
> -	unsigned long flags;
>   	struct occ_response *resp = &occ->resp;
>   	struct p9_sbe_occ *ctx = to_p9_sbe_occ(occ);
>
> -	spin_lock_irqsave(&ctx->client_lock, flags);
> +	mutex_lock(&ctx->client_lock);
>   	if (ctx->sbe)
>   		ctx->client = occ_drv_open(ctx->sbe, 0);
> -	spin_unlock_irqrestore(&ctx->client_lock, flags);
> +	mutex_unlock(&ctx->client_lock);
>
>   	if (!ctx->client)
>   		return -ENODEV;
> @@ -115,7 +113,7 @@ static int p9_sbe_occ_probe(struct platform_device *pdev)
>   	if (!ctx)
>   		return -ENOMEM;
>
> -	spin_lock_init(&ctx->lock);
> +	mutex_init(&ctx->client_lock);
>   	ctx->sbe = pdev->dev.parent;
>   	occ = &ctx->occ;
>   	occ->bus_dev = &pdev->dev;