2020-09-29 20:35:18

by Evan Green

[permalink] [raw]
Subject: [RESEND PATCH] soc: qcom: smp2p: Safely acquire spinlock without IRQs

smp2p_update_bits() should disable interrupts when it acquires its
spinlock. This is important because without the _irqsave, a priority
inversion can occur.

This function is called both with interrupts enabled in
qcom_q6v5_request_stop(), and with interrupts disabled in
ipa_smp2p_panic_notifier(). IRQ handling of spinlocks should be
consistent to avoid the panic notifier deadlocking because it's
sitting on the thread that's already got the lock via _request_stop().

Found via lockdep.

Fixes: 50e99641413e7 ("soc: qcom: smp2p: Qualcomm Shared Memory Point to Point")
Signed-off-by: Evan Green <[email protected]>
---

drivers/soc/qcom/smp2p.c | 5 +++--
1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/drivers/soc/qcom/smp2p.c b/drivers/soc/qcom/smp2p.c
index 07183d731d747..a9709aae54abb 100644
--- a/drivers/soc/qcom/smp2p.c
+++ b/drivers/soc/qcom/smp2p.c
@@ -318,15 +318,16 @@ static int qcom_smp2p_inbound_entry(struct qcom_smp2p *smp2p,
static int smp2p_update_bits(void *data, u32 mask, u32 value)
{
struct smp2p_entry *entry = data;
+ unsigned long flags;
u32 orig;
u32 val;

- spin_lock(&entry->lock);
+ spin_lock_irqsave(&entry->lock, flags);
val = orig = readl(entry->value);
val &= ~mask;
val |= value;
writel(val, entry->value);
- spin_unlock(&entry->lock);
+ spin_unlock_irqrestore(&entry->lock, flags);

if (val != orig)
qcom_smp2p_kick(entry->smp2p);
--
2.26.2


2020-10-16 17:22:34

by Bjorn Andersson

[permalink] [raw]
Subject: Re: [RESEND PATCH] soc: qcom: smp2p: Safely acquire spinlock without IRQs

On Tue 29 Sep 15:30 CDT 2020, Evan Green wrote:

> smp2p_update_bits() should disable interrupts when it acquires its
> spinlock. This is important because without the _irqsave, a priority
> inversion can occur.
>
> This function is called both with interrupts enabled in
> qcom_q6v5_request_stop(), and with interrupts disabled in
> ipa_smp2p_panic_notifier(). IRQ handling of spinlocks should be
> consistent to avoid the panic notifier deadlocking because it's
> sitting on the thread that's already got the lock via _request_stop().
>
> Found via lockdep.
>
> Fixes: 50e99641413e7 ("soc: qcom: smp2p: Qualcomm Shared Memory Point to Point")
> Signed-off-by: Evan Green <[email protected]>

Reviewed-by: Bjorn Andersson <[email protected]>

Regards,
Bjorn

> ---
>
> drivers/soc/qcom/smp2p.c | 5 +++--
> 1 file changed, 3 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/soc/qcom/smp2p.c b/drivers/soc/qcom/smp2p.c
> index 07183d731d747..a9709aae54abb 100644
> --- a/drivers/soc/qcom/smp2p.c
> +++ b/drivers/soc/qcom/smp2p.c
> @@ -318,15 +318,16 @@ static int qcom_smp2p_inbound_entry(struct qcom_smp2p *smp2p,
> static int smp2p_update_bits(void *data, u32 mask, u32 value)
> {
> struct smp2p_entry *entry = data;
> + unsigned long flags;
> u32 orig;
> u32 val;
>
> - spin_lock(&entry->lock);
> + spin_lock_irqsave(&entry->lock, flags);
> val = orig = readl(entry->value);
> val &= ~mask;
> val |= value;
> writel(val, entry->value);
> - spin_unlock(&entry->lock);
> + spin_unlock_irqrestore(&entry->lock, flags);
>
> if (val != orig)
> qcom_smp2p_kick(entry->smp2p);
> --
> 2.26.2
>

2020-10-16 19:00:04

by Stephen Boyd

[permalink] [raw]
Subject: Re: [RESEND PATCH] soc: qcom: smp2p: Safely acquire spinlock without IRQs

Quoting Evan Green (2020-09-29 13:30:57)
> smp2p_update_bits() should disable interrupts when it acquires its
> spinlock. This is important because without the _irqsave, a priority
> inversion can occur.
>
> This function is called both with interrupts enabled in
> qcom_q6v5_request_stop(), and with interrupts disabled in
> ipa_smp2p_panic_notifier(). IRQ handling of spinlocks should be
> consistent to avoid the panic notifier deadlocking because it's
> sitting on the thread that's already got the lock via _request_stop().
>
> Found via lockdep.
>
> Fixes: 50e99641413e7 ("soc: qcom: smp2p: Qualcomm Shared Memory Point to Point")
> Signed-off-by: Evan Green <[email protected]>
> ---

Reviewed-by: Stephen Boyd <[email protected]>