Re: [PATCH v2] mmc: inline the first mmc_scan() on mmc_start_host()

[Dennis Zhou <dennis@kernel.org>]

On Fri, Apr 07, 2023 at 01:24:15AM -0700, Dennis Zhou wrote:
> On Mon, Apr 03, 2023 at 11:50:41AM +0200, Ulf Hansson wrote:
> > On Fri, 31 Mar 2023 at 20:23, Dennis Zhou <dennis@kernel.org> wrote:
> > >
> > > Hi Ulf,
> > >
> > > On Fri, Mar 31, 2023 at 02:43:10PM +0200, Ulf Hansson wrote:
> > > > On Thu, 30 Mar 2023 at 01:48, Dennis Zhou <dennis@kernel.org> wrote:
> > > > >
> > > > > When using dm-verity with a data partition on an emmc device, dm-verity
> > > > > races with the discovery of attached emmc devices. This is because mmc's
> > > > > probing code sets up the host data structure then a work item is
> > > > > scheduled to do discovery afterwards. To prevent this race on init,
> > > > > let's inline the first call to detection, __mm_scan(), and let
> > > > > subsequent detect calls be handled via the workqueue.
> > > >
> > > > In principle, I don't mind the changes in $subject patch, as long as
> > > > it doesn't hurt the overall initialization/boot time. Especially, we
> > > > may have more than one mmc-slot being used, so this needs to be well
> > > > tested.
> > > >
> > >
> > > I unfortunately don't have a device with multiple mmcs available. Is
> > > this something you could help me with?
> > 
> > Yes, I can help to test. Allow me a few days to see what I can do.
> > 
> > Note that, just having one eMMC and one SD card should work too. It
> > doesn't have to be multiple eMMCs.
> > 
> > >
> > > > Although, more importantly, I fail to understand how this is going to
> > > > solve the race condition. Any I/O request to an eMMC or SD requires
> > > > the mmc block device driver to be up and running too, which is getting
> > > > probed from a separate module/driver that's not part of mmc_rescan().
> > >
> > > I believe the call chain is something like this:
> > >
> > > __mmc_rescan()
> > >     mmc_rescan_try_freq()
> > >         mmc_attach_mmc()
> > >             mmc_add_card()
> > >                 device_add()
> > >                     bus_probe_device()
> > >                         mmc_blk_probe()
> > >
> > > The initial calling of this is the host probe. So effectively if there
> > > is a card attached, we're inlining the device_add() call for the card
> > > attached rather than waiting for the workqueue item to kick off.
> > >
> > > dm is a part of late_initcall() while mmc is a module_init(), when built
> > > in becoming a device_initcall(). So this solves a race via the initcall
> > > chain. In the current state, device_initcall() finishes and we move onto
> > > the late_initcall() phase. But now, dm is racing with the workqueue to
> > > init the attached emmc device.
> > 
> > You certainly have a point!
> > 
> > This should work when the mmc blk module is built-in. Even if that
> > doesn't solve the entire problem, it should be a step in the right
> > direction.
> > 
> > I will give it some more thinking and run some tests at my side, then
> > I will get back to you again.
> > 
> 
> Hi Ulf, is there an update on testing with this patch?
> 

Ping.

Thanks,
Dennis