Re: [PATCH] soc: mediatek: cmdq: avoid racing condition with mutex

From: CK Hu
Date: Mon Dec 30 2019 - 01:44:53 EST


Hi, Bibby:

On Thu, 2019-11-21 at 15:29 +0800, Bibby Hsieh wrote:
> If cmdq client is multi thread user, racing will occur without mutex
> protection. It will make the C message queued in mailbox's queue
> always need D message's triggering.
>
> Thread A Thread B Thread C Thread D...
> -----------------------------------------------------------------------------------
> mbox_send_message()
> send_data()
> mbox_send_message()
> *exit
> mbox_send_message()
> *exit
> mbox_client_txdone()
> tx_tick()
> mbox_client_txdone()
> tx_tick()
> mbox_client_txdone()
> tx_tick()
> msg_submit()
> send_data()
> msg_submit()
> *exit
> msg_submit()
> *exit
> -----------------------------------------------------------------------------------
>
> Signed-off-by: Bibby Hsieh <bibby.hsieh@xxxxxxxxxxxx>
> ---
> drivers/soc/mediatek/mtk-cmdq-helper.c | 3 +++
> include/linux/soc/mediatek/mtk-cmdq.h | 1 +
> 2 files changed, 4 insertions(+)
>
> diff --git a/drivers/soc/mediatek/mtk-cmdq-helper.c b/drivers/soc/mediatek/mtk-cmdq-helper.c
> index 9add0fd5fa6c..9e35e0beffaa 100644
> --- a/drivers/soc/mediatek/mtk-cmdq-helper.c
> +++ b/drivers/soc/mediatek/mtk-cmdq-helper.c
> @@ -81,6 +81,7 @@ struct cmdq_client *cmdq_mbox_create(struct device *dev, int index, u32 timeout)
> client->client.dev = dev;
> client->client.tx_block = false;
> client->chan = mbox_request_channel(&client->client, index);
> + mutex_init(&client->mutex);
>
> if (IS_ERR(client->chan)) {
> long err;
> @@ -352,9 +353,11 @@ int cmdq_pkt_flush_async(struct cmdq_pkt *pkt, cmdq_async_flush_cb cb,
> spin_unlock_irqrestore(&client->lock, flags);
> }
>
> + mutex_lock(&client->mutex);
> mbox_send_message(client->chan, pkt);
> /* We can send next packet immediately, so just call txdone. */
> mbox_client_txdone(client->chan, 0);
> + mutex_unlock(&client->mutex);


In [1], Mediatek DRM is the first client to use cmdq and it already has
its own mutex to protect this. I think helper is something help many
user but now I just see Mediatek MDP [2] need this. For DRM, there are
so many useless code in cmdq_pkt_flush_async(). DRM does not need the
timer to check timeout. DRM could do dma_sync_single_for_cpu() in its
callback and need not to create a intermediate callback to do this. I
would agree this patch only when I see two or more user need this.


[1]
https://github.com/ckhu-mediatek/linux.git-tags/commit/4df12ed1866d1104f631e06218bd15fde512a79e
[2] https://patchwork.kernel.org/patch/10945609/

Regards,
CK

>
> return 0;
> }
> diff --git a/include/linux/soc/mediatek/mtk-cmdq.h b/include/linux/soc/mediatek/mtk-cmdq.h
> index a74c1d5acdf3..0f9071cd1bc7 100644
> --- a/include/linux/soc/mediatek/mtk-cmdq.h
> +++ b/include/linux/soc/mediatek/mtk-cmdq.h
> @@ -28,6 +28,7 @@ struct cmdq_client {
> struct mbox_chan *chan;
> struct timer_list timer;
> u32 timeout_ms; /* in unit of microsecond */
> + struct mutex mutex;
> };
>
> /**