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

From: Greg KH
Date: Fri Jun 09 2023 - 02:20:04 EST


On Thu, Jun 08, 2023 at 01:49:00PM -0700, Dennis Zhou wrote:
> On Fri, May 12, 2023 at 01:42:51PM +0200, Ulf Hansson wrote:
> > + Linus,
> >
> > Hi Dennis,
> >
> > On Mon, 3 Apr 2023 at 11:50, Ulf Hansson <ulf.hansson@xxxxxxxxxx> wrote:
> > >
> > > On Fri, 31 Mar 2023 at 20:23, Dennis Zhou <dennis@xxxxxxxxxx> 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@xxxxxxxxxx> 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.
> > >
> > > Kind regards
> > > Uffe
> > >
> > > > >
> > > > > >
> > > > > > Signed-off-by: Dennis Zhou <dennis@xxxxxxxxxx>
> > > > > > ---
> > > > > > Sigh.. fix missing static declaration.
> > > > > >
> > > > > > drivers/mmc/core/core.c | 15 +++++++++++----
> > > > > > 1 file changed, 11 insertions(+), 4 deletions(-)
> > > > > >
> > > > > > diff --git a/drivers/mmc/core/core.c b/drivers/mmc/core/core.c
> > > > > > index 368f10405e13..fda7ee57dee3 100644
> > > > > > --- a/drivers/mmc/core/core.c
> > > > > > +++ b/drivers/mmc/core/core.c
> > > > > > @@ -2185,10 +2185,8 @@ int mmc_card_alternative_gpt_sector(struct mmc_card *card, sector_t *gpt_sector)
> > > > > > }
> > > > > > EXPORT_SYMBOL(mmc_card_alternative_gpt_sector);
> > > > > >
> > > > > > -void mmc_rescan(struct work_struct *work)
> > > > > > +static void __mmc_rescan(struct mmc_host *host)
> > > > > > {
> > > > > > - struct mmc_host *host =
> > > > > > - container_of(work, struct mmc_host, detect.work);
> > > > > > int i;
> > > > > >
> > > > > > if (host->rescan_disable)
> > > > > > @@ -2249,6 +2247,14 @@ void mmc_rescan(struct work_struct *work)
> > > > > > mmc_schedule_delayed_work(&host->detect, HZ);
> > > > > > }
> > > > > >
> > > > > > +void mmc_rescan(struct work_struct *work)
> > > > > > +{
> > > > > > + struct mmc_host *host =
> > > > > > + container_of(work, struct mmc_host, detect.work);
> > > > > > +
> > > > > > + __mmc_rescan(host);
> > > > > > +}
> > > > > > +
> > > > > > void mmc_start_host(struct mmc_host *host)
> > > > > > {
> > > > > > host->f_init = max(min(freqs[0], host->f_max), host->f_min);
> > > > > > @@ -2261,7 +2267,8 @@ void mmc_start_host(struct mmc_host *host)
> > > > > > }
> > > > > >
> > > > > > mmc_gpiod_request_cd_irq(host);
> > > > > > - _mmc_detect_change(host, 0, false);
> > > > > > + host->detect_change = 1;
> > > > > > + __mmc_rescan(host);
> > > > > > }
> > > > > >
> > > > > > void __mmc_stop_host(struct mmc_host *host)
> > > > > > --
> > > > > > 2.40.0
> > > > > >
> >
> > My apologies for the long delay. I finally managed to test this.
> >
> > I decided to pick an old arm32 based platform. An ST-Ericsson HREF,
> > based upon the ux500 SoC. It's quite good to use for these types of
> > tests as it has two eMMCs soldered, an embedded SDIO (for WiFi) and an
> > SD-card slot. So in total there are 4 devices that get probed.
> >
> > The SDIO card isn't detected properly, but always fails in the similar
> > way (thus I left it out from the below data). I tested both with and
> > without an SD card inserted during boot, to get some more data to
> > compare. These are the summary from my tests:
> >
> > v6.4-rc1 without SD card:
> > ~2.18s - MMC1 (eMMC)
> > ~3.33s - MMC3 (eMMC)
> > ~5.91s - kernel boot complete
> >
> > v6.4-rc1 with an SD card:
> > ~2.18s - MMC1 (eMMC)
> > ~3.45s - MMC3 (eMMC)
> > ~3.57s - MMC2 (SD)
> > ~5.76s - kernel boot complete
> >
> > v6.4-rc1 + patch without SD card:
> > ~2.24s - MMC1 (eMMC)
> > ~3.58s - MMC3 (eMMC)
> > ~5.96s - kernel boot complete
> >
> > v6.4-rc1 + patch with an SD card:
> > ~2.24s - MMC1 (eMMC)
> > ~3.73s - MMC2 (SD)
> > ~3.98s - MMC3 (eMMC)
> > ~6.73s - kernel boot complete
> >
> > By looking at these results, I was kind of surprised. I was thinking
> > that the asynchronous probe should address the parallelism problem.
> > Then I discovered that it in fact, hasn't been enabled for the mmci
> > driver that is being used for this platform. Huh, I was under the
> > assumption that it has been enabled for all mmc hosts by now. :-)
> >
> > Okay, so I am going to run another round of tests, with async probe
> > enabled for the mmci driver too. I will let you know the results as
> > soon as I can.
> >
> > Kind regards
> > Uffe
>
> Hi Uffe,
>
> Kindly this has been way too long for review. It's been over 3 months.
> What's going on here?
>
> I think there's a misunderstanding too. Without this fix, the machine
> doesn't even boot. I'm not sure why perf is the blocking question here.

Well you can not degrade performance of existing machines that work
today, right? That would be a regression and it seems that you are
doing that if I read the numbers above correctly.

> Greg, is there another tree I can run this through?

Why would you want to route around a maintainer just to get a patch that
would have to be reverted applied? :)

thanks,

greg k-h