On Tue, 6 Jul 2021 at 11:09, Christian Löhle <CLoehle@xxxxxxxxxxxxxx> wrote:
>
> Hey Uffe,
>
> >> +static int is_return_to_tran_cmd(struct mmc_command *cmd)
> >> +{
> >> + /*
> >> + * Cards will never return to TRAN after completing
> >> + * identification commands or MMC_SEND_STATUS if they are not selected.
> >> + */
> >> + switch (cmd->opcode) {
> >> + case MMC_GO_IDLE_STATE:
> >> + case MMC_SEND_OP_COND:
> >> + case MMC_ALL_SEND_CID:
> >> + case MMC_SET_RELATIVE_ADDR:
> >> + case MMC_SET_DSR:
> >> + case MMC_SLEEP_AWAKE:
> >> + case MMC_SELECT_CARD:
> >> + case MMC_SEND_CSD:
> >> + case MMC_SEND_CID:
> >> + case MMC_SEND_STATUS:
> >> + case MMC_GO_INACTIVE_STATE:
> >> + case MMC_APP_CMD:
> >> + return false;
> >> + default:
> >> + return true;
> >> + }
> >> +}
> >>
> >What exactly are you trying to do with the user space program through
> >the mmc ioctl with all these commands? The mmc ioctl interface is not
> >designed to be used like that.
> >
> >In principle, it looks like we should support a complete
> >re-initialization of the card. I am sorry, but no thanks! This doesn't
> >work, but more importantly, this should be managed solely by the
> >kernel, in my opinion.
>
> Doing initialization itself through ioctl is silly, I agree, and does
> not work on other ends. This patch is not about that. it just explicitly disables
> any CMD13 polling for TRAN for some of those commands. the behavior
> for such commands thus is the same as without the patch.
You are right.
But, what I think is bothering me with the approach, is that it looks
like we are starting to add special treatment of a whole bunch of
different commands.
> The reason for this patch is to not run into the race condition that a
> following (ioctl) command will be rejected because the card is in e.g. PROG state
> of a previous ioctl command. As stated earlier, I encountered this a lot when
> doing a unlock force erase -> lock/set, in both scenarios, issued as two single
> ioctl commands and bundled together.
I understand. I would rather see a patch that adds support, explicitly
for this case.
> But this race condition exists on any (non-R1b/ RPBM) currently. As there is
> no CMD13 polling happening after the response (or whenever the driver marks
> the request as done), the card's status is therefore generally unknown.
So the commands to unlock/lock, etc, don't have R1B, but can still
cause the card to become busy after the response has been delivered,
right?
As I said, then please add this as an explicit check to do polling,
then I would be happy. :-)
>
> So in short I don;t want to do anything too crazy from userspace, but the
> alternative now is to do like 100ms sleeps in the hope that the card is
> actually finished with the issued command (not just the host driver so to say).
Yeah, that sounds suboptimal, we can do better than that.
Kind regards
Uffe
