Re: [PATCH V2] mmc: sdhci: don't read cd-gpio while holding spinlock

[Adrian Hunter <adrian.hunter@intel.com>]

On 24.03.2014 18:18, Andrew Bresticker wrote:
> On Sun, Mar 23, 2014 at 11:19 PM, Adrian Hunter <adrian.hunter@intel.com> wrote:
>> On 24.03.2014 03:43, Andrew Bresticker wrote:
>>> sdhci_request() and sdhci_card_event() read the cd-gpio via
>>> mmc_gpio_get_cd(), which can sleep, while holding host->lock.
>>> This may result in the following BUG:
>>>
>>>   BUG: spinlock wrong CPU on CPU#2, kworker/u8:16/4296
>>>   lock: 0xea6b9c80, .magic: dead4ead, .owner: kworker/u8:16/4296, .owner_cpu: 0
>>>   CPU: 2 PID: 4296 Comm: kworker/u8:16 Tainted: G         C   3.10.18 #137
>>>   Workqueue: kmmcd mmc_rescan
>>>   [<8020cf8c>] (unwind_backtrace+0x0/0x118) from [<8020a0c8>] (show_stack+0x20/0x24)
>>>   [<8020a0c8>] (show_stack+0x20/0x24) from [<8075e5b8>] (dump_stack+0x20/0x28)
>>>   [<8075e5b8>] (dump_stack+0x20/0x28) from [<804184a8>] (spin_dump+0x80/0x94)
>>>   [<804184a8>] (spin_dump+0x80/0x94) from [<804184e8>] (spin_bug+0x2c/0x30)
>>>   [<804184e8>] (spin_bug+0x2c/0x30) from [<80418790>] (do_raw_spin_unlock+0x94/0xd4)
>>>   [<80418790>] (do_raw_spin_unlock+0x94/0xd4) from [<80761a44>] (_raw_spin_unlock_irqrestore+0x1c/0x24)
>>>   [<80761a44>] (_raw_spin_unlock_irqrestore+0x1c/0x24) from [<805ff66c>] (sdhci_request+0x1c8/0x1d0)
>>>   [<805ff66c>] (sdhci_request+0x1c8/0x1d0) from [<805ebb5c>] (mmc_start_request+0xec/0xf4)
>>>   [<805ebb5c>] (mmc_start_request+0xec/0xf4) from [<805ebcbc>] (mmc_wait_for_req+0x80/0xf4)
>>>   ...
>>>
>>> Read the cd-gpio before acquiring the spinlock instead.
>>>
>>> Signed-off-by: Andrew Bresticker <abrestic@chromium.org>
>>> ---
>>>  drivers/mmc/host/sdhci.c | 7 ++++---
>>>  1 file changed, 4 insertions(+), 3 deletions(-)
>>>
>>> diff --git a/drivers/mmc/host/sdhci.c b/drivers/mmc/host/sdhci.c
>>> index 04a5e25..328b250 100644
>>> --- a/drivers/mmc/host/sdhci.c
>>> +++ b/drivers/mmc/host/sdhci.c
>>> @@ -1340,6 +1340,7 @@ static void sdhci_request(struct mmc_host *mmc, struct mmc_request *mrq)
>>>       u32 tuning_opcode;
>>>
>>>       host = mmc_priv(mmc);
>>> +     present = mmc_gpio_get_cd(host->mmc);
>>>
>>>       sdhci_runtime_pm_get(host);
>>>
>>> @@ -1371,7 +1372,6 @@ static void sdhci_request(struct mmc_host *mmc, struct mmc_request *mrq)
>>>        *     zero: cd-gpio is used, and card is removed
>>>        *     one: cd-gpio is used, and card is present
>>>        */
>>> -     present = mmc_gpio_get_cd(host->mmc);
>>>       if (present < 0) {
>>>               /* If polling, assume that the card is always present. */
>>>               if (host->quirks & SDHCI_QUIRK_BROKEN_CARD_DETECTION)
>>> @@ -2082,15 +2082,16 @@ static void sdhci_card_event(struct mmc_host *mmc)
>>>  {
>>>       struct sdhci_host *host = mmc_priv(mmc);
>>>       unsigned long flags;
>>> +     int present;
>>>
>>>       /* First check if client has provided their own card event */
>>>       if (host->ops->card_event)
>>>               host->ops->card_event(host);
>>>
>>> +     present = sdhci_get_cd(mmc);
>>
>> Shouldn't that be sdhci_do_get_cd(mmc) ?
> 
> No I do mean shdic_get_cd().  We might be runtime suspended so we need
> sdhci_get_cd() to increment the device usage count and do a runtime
> resume if necessary.

Very true.

Acked-by: Adrian Hunter <adrian.hunter@intel.com>

> 
>>
>>>       spin_lock_irqsave(&host->lock, flags);
>>>
>>> -     /* Check host->mrq first in case we are runtime suspended */
>>> -     if (host->mrq && !sdhci_do_get_cd(host)) {
>>> +     if (host->mrq && !present) {
>>>               pr_err("%s: Card removed during transfer!\n",
>>>                       mmc_hostname(host->mmc));
>>>               pr_err("%s: Resetting controller.\n",
>>>
>>
> 
>