With the to-be-fixed commit, the reset_work handler cleared 'host->mrq' outside of the spinlock protected critical section. That leaves a small race window during execution of 'tmio_mmc_reset()' where the done_work handler could grab a pointer to the now invalid 'host->mrq'. Both would use it to call mmc_request_done() causing problems (see Link). However, 'host->mrq' cannot simply be cleared earlier inside the critical section. That would allow new mrqs to come in asynchronously while the actual reset of the controller still needs to be done. So, like 'tmio_mmc_set_ios()', an ERR_PTR is used to prevent new mrqs from coming in but still avoiding concurrency between work handlers. Reported-by: Dirk Behme <dirk.behme@xxxxxxxxxxxx> Closes: https://lore.kernel.org/all/20240220061356.3001761-1-dirk.behme@xxxxxxxxxxxx/ Signed-off-by: Wolfram Sang <wsa+renesas@xxxxxxxxxxxxxxxxxxxx> Fixes: df3ef2d3c92c ("mmc: protect the tmio_mmc driver against a theoretical race") --- Dirk: could you get this tested on your affected setups? I am somewhat optimistic that this is already enough. For sure, it is a needed first step. drivers/mmc/host/tmio_mmc_core.c | 2 ++ 1 file changed, 2 insertions(+) diff --git a/drivers/mmc/host/tmio_mmc_core.c b/drivers/mmc/host/tmio_mmc_core.c index be7f18fd4836..c253d176db69 100644 --- a/drivers/mmc/host/tmio_mmc_core.c +++ b/drivers/mmc/host/tmio_mmc_core.c @@ -259,6 +259,8 @@ static void tmio_mmc_reset_work(struct work_struct *work) else mrq->cmd->error = -ETIMEDOUT; + /* No new calls yet, but disallow concurrent tmio_mmc_done_work() */ + host->mrq = ERR_PTR(-EBUSY); host->cmd = NULL; host->data = NULL; -- 2.43.0
