Re: [PATCH v2] [RESEND] mmc: add Toshiba PCI SD controller driver

[Ondrej Zary <linux@rainbow-software.org>]

On Thursday 06 November 2014 00:02:24 Ondrej Zary wrote:
> On Tuesday 04 November 2014 22:41:44 Ondrej Zary wrote:
> > On Tuesday 04 November 2014 12:09:44 Ulf Hansson wrote:
> > > On 2 November 2014 22:51, Ondrej Zary <linux@rainbow-software.org> wrote:
> > > > This patch resurrects an old never-finished driver for Toshiba PCI SD
> > > > controllers found in some older Toshiba laptops (such as Portege R100):
> > > >
> > > > 02:0d.0 System peripheral [0880]: Toshiba America Info Systems SD TypA Controller [1179:0805] (rev 05)
> > > >
> > > > The code is fixed, cleaned up and successfully tested with SD, SDHC, SDXC and
> > > > MMC cards on Portege R100. (MMC cards don't even work in Windows!)
> > > > SDIO probably does not work (don't have any SDIO card).
> > > >
> > > > The hardware is slow (around 2 MB/s - same in Windows) because it does not
> > > > support bus mastering (busmaster enable bit cannot be set in PCI control reg).
> > > > Also the card clock is limited to 16MHz (33MHz PCI clock divided by 2).
> > > >
> > > > Signed-off-by: Ondrej Zary <linux@rainbow-software.org>
> > > 
> > > Hi Ondrej,
> > > 
> > > Sorry for a very very late reply.
> > > 
> > 
> > ...
> > 
> > > > +static void toshsd_init(struct toshsd_host *host);
> > > > +static void toshsd_set_ios_unlocked(struct mmc_host *mmc, struct mmc_ios *ios);
> > > 
> > > I would implement these functions at the proper place instead of
> > > having them defined here.
> > > 
> > > Moreover I think toshsd_set_ios_unlocked() could be renamed to
> > > "__toshsd_set_ios()".
> > 
> > OK, will do.
> >  
> > > > +
> > > > +static inline u16 toshsd_readw(struct toshsd_host *host, u16 reg)
> > > > +{
> > > > +       return ioread16(host->ioaddr + reg);
> > > > +}
> > > > +
> > > > +static inline u32 toshsd_readl(struct toshsd_host *host, u16 reg)
> > > > +{
> > > > +       return ioread32(host->ioaddr + reg);
> > > > +}
> > > > +
> > > > +static inline void toshsd_writew(struct toshsd_host *host, u16 reg, u16 val)
> > > > +{
> > > > +       iowrite16(val, host->ioaddr + reg);
> > > > +}
> > > > +
> > > > +static inline void toshsd_writel(struct toshsd_host *host, u16 reg, u32 val)
> > > > +{
> > > > +       iowrite32(val, host->ioaddr + reg);
> > > > +}
> > > > +
> > > > +static inline void toshsd_readl_rep(struct toshsd_host *host, u16 reg,
> > > > +                                   void *dst, unsigned long count)
> > > > +{
> > > > +       ioread32_rep(host->ioaddr + reg, dst, count);
> > > > +}
> > > > +
> > > > +static inline void toshsd_writel_rep(struct toshsd_host *host, u16 reg,
> > > > +                                    const void *src, unsigned long count)
> > > > +{
> > > > +       iowrite32_rep(host->ioaddr + reg, src, count);
> > > > +}
> > > 
> > > To me, all these wrapper functions seems a bit ugly. How about
> > > invoking io* functions directly instead?
> > 
> > It's a matter of preference, some drivers use wrappers, some don't.
> > I can remove them if they're not recommended in mmc subsystem.
> >  
> > ...
> > 
> > > > +               tasklet_schedule(&host->data_read_tasklet);
> > > 
> > > Instead of using a tasklet, I would advise to use a threaded IRQ.
> >  
> > Haven't used threaded IRQ yet, will try.
> > 
> > ...
> > 
> > > > +#ifdef CONFIG_PM
> > > 
> > > This should be CONFIG_PM_SLEEP.
> > > 
> > > > +
> > > > +static int toshsd_suspend(struct pci_dev *pdev, pm_message_t state)
> > > 
> > > This is the legacy version of system PM callbacks. You need to convert
> > > to the modern ones instead.
> > > 
> > > > +{
> > > > +       struct toshsd_host *host = pci_get_drvdata(pdev);
> > > > +
> > > > +       toshsd_powerdown(host);
> > > > +
> > > > +       pci_save_state(pdev);
> > > > +       pci_enable_wake(pdev, pci_choose_state(pdev, state), 0);
> > > > +       pci_disable_device(pdev);
> > > > +       pci_set_power_state(pdev, pci_choose_state(pdev, state));
> > > > +
> > > > +       return 0;
> > > > +}
> > > > +
> > > > +static int toshsd_resume(struct pci_dev *pdev)
> > > 
> > > This is the legacy version of system PM callbacks. You need to convert
> > > to the modern ones instead.
> > 
> > I just converted them and found that suspend does not work on current kernels
> > when a SD card is inserted (even with unmodified toshsd driver):
> > [  188.960862] dpm_run_callback(): mmc_bus_suspend+0x0/0x2c [mmc_core] returns -110
> > [  188.960867] PM: Device mmc0:b368 failed to suspend: error -110
> > [  188.960869] PM: Some devices failed to suspend, or early wake event detected
> > 
> > Is it a kernel bug or the driver is missing something?
> > 
> 
> The same kernel with sdhci-pci works fine (on another HW) so the problem is
> limited only to toshsd...
> Seems that the MMC core is trying to do something with the card and it fails
> with timeout:
> [  885.750001] PM: Entering mem sleep
> [  885.750023] Suspending console(s) (use no_console_suspend to debug)
> [  885.750281] Command opcode: 7
> [  885.750432] timeout!!!!!
> [  885.750453] Command opcode: 7
> [  885.750597] timeout!!!!!
> [  885.750615] Command opcode: 7
> [  885.750758] timeout!!!!!
> [  885.750776] Command opcode: 7
> [  885.750920] timeout!!!!!
> [  885.750956] dpm_run_callback(): mmc_bus_suspend+0x0/0x2c [mmc_core] returns -110
> [  885.750961] PM: Device mmc0:0007 failed to suspend: error -110
> 

Found a bug in the header file, probably causing all no-response commands to fail:
#define SD_CMD_RESP_TYPE_NORMAL                (0 << 8)

Both the name and value were wrong, it should be:
#define SD_CMD_RESP_TYPE_NONE          (3 << 8)

Older kernels were probably not sending MMC_SELECT_CARD on suspend.

-- 
Ondrej Zary