When the `CONFIG_I2C_SLAVE` option is enabled and the device operates
as a slave, a situation arises where the master sends a START signal
without the accompanying STOP signal. This action results in a
persistent I2C bus timeout. The core issue stems from the fact that
the i2c controller remains in a slave read state without a timeout
mechanism. As a consequence, the bus perpetually experiences timeouts.
In this case, the i2c bus will be reset, but the slave_state reset is
missing.
Fixes: fee465150b45 ("i2c: aspeed: Reset the i2c controller when timeout occurs")
Signed-off-by: Jian Zhang <[email protected]>
---
drivers/i2c/busses/i2c-aspeed.c | 1 +
1 file changed, 1 insertion(+)
diff --git a/drivers/i2c/busses/i2c-aspeed.c b/drivers/i2c/busses/i2c-aspeed.c
index 5a416b39b818..18f618625472 100644
--- a/drivers/i2c/busses/i2c-aspeed.c
+++ b/drivers/i2c/busses/i2c-aspeed.c
@@ -933,6 +933,7 @@ static int aspeed_i2c_init(struct aspeed_i2c_bus *bus,
/* If slave has already been registered, re-enable it. */
if (bus->slave)
__aspeed_i2c_reg_slave(bus, bus->slave->addr);
+ bus->slave_state = ASPEED_I2C_SLAVE_INACTIVE;
#endif /* CONFIG_I2C_SLAVE */
/* Set interrupt generation of I2C controller */
--
2.30.2
Hi Jian,
On Wed, Sep 27, 2023 at 11:42:43PM +0800, Jian Zhang wrote:
> When the `CONFIG_I2C_SLAVE` option is enabled and the device operates
> as a slave, a situation arises where the master sends a START signal
> without the accompanying STOP signal. This action results in a
> persistent I2C bus timeout. The core issue stems from the fact that
> the i2c controller remains in a slave read state without a timeout
> mechanism. As a consequence, the bus perpetually experiences timeouts.
>
> In this case, the i2c bus will be reset, but the slave_state reset is
> missing.
>
> Fixes: fee465150b45 ("i2c: aspeed: Reset the i2c controller when timeout occurs")
> Signed-off-by: Jian Zhang <[email protected]>
Why I'm failing to find your v1 patch? And where is the
changelog?
Andi
> From: "Andi Shyti"<[email protected]>
> Date: Thu, Sep 28, 2023, 22:51
> Subject: [External] Re: [PATCH v2] i2c: aspeed: Fix i2c bus hang in slave read
> To: "Jian Zhang"<[email protected]>
> Cc: <[email protected]>, <[email protected]>, <[email protected]>, <[email protected]>, <[email protected]>, <[email protected]>, <[email protected]>, "Tommy Huang"<[email protected]>, "Wolfram Sang"<[email protected]>, "open list:ARM/ASPEED I2C DRIVER"<[email protected]>, "moderated list:ARM/ASPEED I2C DRIVER"<[email protected]>, "moderated list:ARM/ASPEED MACHINE SUPPORT"<[email protected]>, "moderated list:ARM/ASPEED MACHINE SUPPORT"<[email protected]>, "open list"<[email protected]>
> Hi Jian,
>
> On Wed, Sep 27, 2023 at 11:42:43PM +0800, Jian Zhang wrote:
> > When the `CONFIG_I2C_SLAVE` option is enabled and the device operates
> > as a slave, a situation arises where the master sends a START signal
> > without the accompanying STOP signal. This action results in a
> > persistent I2C bus timeout. The core issue stems from the fact that
> > the i2c controller remains in a slave read state without a timeout
> > mechanism. As a consequence, the bus perpetually experiences timeouts.
> >
> > In this case, the i2c bus will be reset, but the slave_state reset is
> > missing.
> >
> > Fixes: fee465150b45 ("i2c: aspeed: Reset the i2c controller when timeout occurs")
> > Signed-off-by: Jian Zhang <[email protected]>
>
> Why I'm failing to find your v1 patch? And where is the
> changelog?
Sorry, something was missing,
v2:
* remove the i2c slave reset and only move the `bus->slave_state =
ASPEED_I2C_SLAVE_INACTIVE` to the aspeed_i2c_init
[0]: https://lore.kernel.org/linux-arm-kernel/[email protected]/T/
Jian
>
> Andi
On Wed, Sep 27, 2023 at 11:42:43PM +0800, Jian Zhang wrote:
> When the `CONFIG_I2C_SLAVE` option is enabled and the device operates
> as a slave, a situation arises where the master sends a START signal
> without the accompanying STOP signal. This action results in a
> persistent I2C bus timeout. The core issue stems from the fact that
> the i2c controller remains in a slave read state without a timeout
> mechanism. As a consequence, the bus perpetually experiences timeouts.
>
> In this case, the i2c bus will be reset, but the slave_state reset is
> missing.
>
> Fixes: fee465150b45 ("i2c: aspeed: Reset the i2c controller when timeout occurs")
> Signed-off-by: Jian Zhang <[email protected]>
Somebody wants to add tags here? I think it should go to my pull request
this week.
Hi Jian,
On Thu, Sep 28, 2023 at 11:04:23AM -0400, Jian Zhang wrote:
> > From: "Andi Shyti"<[email protected]>
> > Date:? Thu, Sep 28, 2023, 22:51
> > Subject:? [External] Re: [PATCH v2] i2c: aspeed: Fix i2c bus hang in slave read
> > To: "Jian Zhang"<[email protected]>
> > Cc: <[email protected]>, <[email protected]>, <[email protected]>, <[email protected]>, <[email protected]>, <[email protected]>, <[email protected]>, "Tommy Huang"<[email protected]>, "Wolfram Sang"<[email protected]>, "open list:ARM/ASPEED I2C DRIVER"<[email protected]>, "moderated list:ARM/ASPEED I2C DRIVER"<[email protected]>, "moderated list:ARM/ASPEED MACHINE SUPPORT"<[email protected]>, "moderated list:ARM/ASPEED MACHINE SUPPORT"<[email protected]>, "open list"<[email protected]>
> > Hi Jian,
> >
> > On Wed, Sep 27, 2023 at 11:42:43PM +0800, Jian Zhang wrote:
> > > When the `CONFIG_I2C_SLAVE` option is enabled and the device operates
> > > as a slave, a situation arises where the master sends a START signal
> > > without the accompanying STOP signal. This action results in a
> > > persistent I2C bus timeout. The core issue stems from the fact that
> > > the i2c controller remains in a slave read state without a timeout
> > > mechanism. As a consequence, the bus perpetually experiences timeouts.
> > >
> > > In this case, the i2c bus will be reset, but the slave_state reset is
> > > missing.
Acked-by: Andi Shyti <[email protected]>
I checked the flow in the driver and makes sense to me. I'd also
love a last minute comment from Brendan or Benjamin or Joel.
> > > Fixes: fee465150b45 ("i2c: aspeed: Reset the i2c controller when timeout occurs")
> > > Signed-off-by: Jian Zhang <[email protected]>
> >
> > Why I'm failing to find your v1 patch? And where is the
> > changelog?
> Sorry, something was missing,
> v2:
> * remove the i2c slave reset and only move the `bus->slave_state =
> ASPEED_I2C_SLAVE_INACTIVE` to the aspeed_i2c_init
>
> [0]: https://lore.kernel.org/linux-arm-kernel/[email protected]/T/
Thanks! I should really check my filters here.
Andi
> Jian
> >
> > Andi
On Fri, 2023-09-29 at 09:39 +0200, Wolfram Sang wrote:
> On Wed, Sep 27, 2023 at 11:42:43PM +0800, Jian Zhang wrote:
> > When the `CONFIG_I2C_SLAVE` option is enabled and the device operates
> > as a slave, a situation arises where the master sends a START signal
> > without the accompanying STOP signal. This action results in a
> > persistent I2C bus timeout. The core issue stems from the fact that
> > the i2c controller remains in a slave read state without a timeout
> > mechanism. As a consequence, the bus perpetually experiences timeouts.
> >
> > In this case, the i2c bus will be reset, but the slave_state reset is
> > missing.
> >
> > Fixes: fee465150b45 ("i2c: aspeed: Reset the i2c controller when timeout occurs")
> > Signed-off-by: Jian Zhang <[email protected]>
>
> Somebody wants to add tags here? I think it should go to my pull request
> this week.
>
I've tested this patch applied on top of fee465150b45 on an AST2600 and
the the system behaviour doesn't seem worse. However, I can still lock
the bus up and trigger a hung task panic by surprise-unplugging things.
I'll poke around to see if I can get to the bottom of that.
Resetting the slave state makes sense, so with the above observation
aside:
Tested-by: Andrew Jeffery <[email protected]>
Reviewed-by: Andrew Jeffery <[email protected]>
That said I do wonder whether we should update the slave state in the
same place we're updating the hardware state. It would cover off the
gap identified by Jian if it were to ever occur anywhere else.
Something like:
diff --git a/drivers/i2c/busses/i2c-aspeed.c b/drivers/i2c/busses/i2c-
aspeed.c
index 5a416b39b818..28e2a5fc4528 100644
--- a/drivers/i2c/busses/i2c-aspeed.c
+++ b/drivers/i2c/busses/i2c-aspeed.c
@@ -749,6 +749,8 @@ static void __aspeed_i2c_reg_slave(struct
aspeed_i2c_bus *bus, u16 slave_addr)
func_ctrl_reg_val = readl(bus->base + ASPEED_I2C_FUN_CTRL_REG);
func_ctrl_reg_val |= ASPEED_I2CD_SLAVE_EN;
writel(func_ctrl_reg_val, bus->base + ASPEED_I2C_FUN_CTRL_REG);
+
+ bus->slave_state = ASPEED_I2C_SLAVE_INACTIVE;
}
static int aspeed_i2c_reg_slave(struct i2c_client *client)
@@ -765,7 +767,6 @@ static int aspeed_i2c_reg_slave(struct i2c_client
*client)
__aspeed_i2c_reg_slave(bus, client->addr);
bus->slave = client;
- bus->slave_state = ASPEED_I2C_SLAVE_INACTIVE;
spin_unlock_irqrestore(&bus->lock, flags);
return 0;
On 04/10/2023 13:08, Andrew Jeffery wrote:
> On Fri, 2023-09-29 at 09:39 +0200, Wolfram Sang wrote:
>> On Wed, Sep 27, 2023 at 11:42:43PM +0800, Jian Zhang wrote:
>>> When the `CONFIG_I2C_SLAVE` option is enabled and the device operates
>>> as a slave, a situation arises where the master sends a START signal
>>> without the accompanying STOP signal. This action results in a
>>> persistent I2C bus timeout. The core issue stems from the fact that
>>> the i2c controller remains in a slave read state without a timeout
>>> mechanism. As a consequence, the bus perpetually experiences timeouts.
>>>
>>> In this case, the i2c bus will be reset, but the slave_state reset is
>>> missing.
>>>
>>> Fixes: fee465150b45 ("i2c: aspeed: Reset the i2c controller when timeout occurs")
>>> Signed-off-by: Jian Zhang <[email protected]>
>>
>> Somebody wants to add tags here? I think it should go to my pull request
>> this week.
>>
>
> I've tested this patch applied on top of fee465150b45 on an AST2600 and
> the the system behaviour doesn't seem worse. However, I can still lock
> the bus up and trigger a hung task panic by surprise-unplugging things.
> I'll poke around to see if I can get to the bottom of that.
>
> Resetting the slave state makes sense, so with the above observation
> aside:
>
> Tested-by: Andrew Jeffery <[email protected]>
> Reviewed-by: Andrew Jeffery <[email protected]>
>
> That said I do wonder whether we should update the slave state in the
> same place we're updating the hardware state. It would cover off the
> gap identified by Jian if it were to ever occur anywhere else.
> Something like:
>
> diff --git a/drivers/i2c/busses/i2c-aspeed.c b/drivers/i2c/busses/i2c-
> aspeed.c
> index 5a416b39b818..28e2a5fc4528 100644
> --- a/drivers/i2c/busses/i2c-aspeed.c
> +++ b/drivers/i2c/busses/i2c-aspeed.c
> @@ -749,6 +749,8 @@ static void __aspeed_i2c_reg_slave(struct
> aspeed_i2c_bus *bus, u16 slave_addr)
> func_ctrl_reg_val = readl(bus->base + ASPEED_I2C_FUN_CTRL_REG);
> func_ctrl_reg_val |= ASPEED_I2CD_SLAVE_EN;
> writel(func_ctrl_reg_val, bus->base + ASPEED_I2C_FUN_CTRL_REG);
> +
> + bus->slave_state = ASPEED_I2C_SLAVE_INACTIVE;
> }
>
> static int aspeed_i2c_reg_slave(struct i2c_client *client)
> @@ -765,7 +767,6 @@ static int aspeed_i2c_reg_slave(struct i2c_client
> *client)
> __aspeed_i2c_reg_slave(bus, client->addr);
>
> bus->slave = client;
> - bus->slave_state = ASPEED_I2C_SLAVE_INACTIVE;
> spin_unlock_irqrestore(&bus->lock, flags);
>
> return 0;
>
>
We tested both Jian's patch and Andrew's patch on our MCTP-i2c bus
(ast2600 based BMC) and see both patches work well.
We currently use upstream i2c-aspeed.c driver with the commit [1]
backported. Without that commit, we frequently experienced the bus hang
(due to bus arbitration) and it is unable to recover.
But, by reverting that commit and with Jian or Andrew's patch, we see
the bus could be able to recover so we think both changes are good.
[1]
https://github.com/AspeedTech-BMC/linux/commit/11a94e5918aa0f87c828d63fd254dd60ab2505e5
Anyway, I would prefer Andrew's way because the bus->slave_state must
always be reset to ASPEED_I2C_SLAVE_INACTIVE everytime
__aspeed_i2c_reg_slave() is called.
Thanks
- Quan
On Thu, 2023-10-05 at 14:55 +0700, Quan Nguyen wrote:
>
> On 04/10/2023 13:08, Andrew Jeffery wrote:
> > On Fri, 2023-09-29 at 09:39 +0200, Wolfram Sang wrote:
> > > On Wed, Sep 27, 2023 at 11:42:43PM +0800, Jian Zhang wrote:
> > > > When the `CONFIG_I2C_SLAVE` option is enabled and the device operates
> > > > as a slave, a situation arises where the master sends a START signal
> > > > without the accompanying STOP signal. This action results in a
> > > > persistent I2C bus timeout. The core issue stems from the fact that
> > > > the i2c controller remains in a slave read state without a timeout
> > > > mechanism. As a consequence, the bus perpetually experiences timeouts.
> > > >
> > > > In this case, the i2c bus will be reset, but the slave_state reset is
> > > > missing.
> > > >
> > > > Fixes: fee465150b45 ("i2c: aspeed: Reset the i2c controller when timeout occurs")
> > > > Signed-off-by: Jian Zhang <[email protected]>
> > >
> > > Somebody wants to add tags here? I think it should go to my pull request
> > > this week.
> > >
> >
> > I've tested this patch applied on top of fee465150b45 on an AST2600 and
> > the the system behaviour doesn't seem worse. However, I can still lock
> > the bus up and trigger a hung task panic by surprise-unplugging things.
> > I'll poke around to see if I can get to the bottom of that.
> >
> > Resetting the slave state makes sense, so with the above observation
> > aside:
> >
> > Tested-by: Andrew Jeffery <[email protected]>
> > Reviewed-by: Andrew Jeffery <[email protected]>
> >
> > That said I do wonder whether we should update the slave state in the
> > same place we're updating the hardware state. It would cover off the
> > gap identified by Jian if it were to ever occur anywhere else.
> > Something like:
> >
> > diff --git a/drivers/i2c/busses/i2c-aspeed.c b/drivers/i2c/busses/i2c-
> > aspeed.c
> > index 5a416b39b818..28e2a5fc4528 100644
> > --- a/drivers/i2c/busses/i2c-aspeed.c
> > +++ b/drivers/i2c/busses/i2c-aspeed.c
> > @@ -749,6 +749,8 @@ static void __aspeed_i2c_reg_slave(struct
> > aspeed_i2c_bus *bus, u16 slave_addr)
> > func_ctrl_reg_val = readl(bus->base + ASPEED_I2C_FUN_CTRL_REG);
> > func_ctrl_reg_val |= ASPEED_I2CD_SLAVE_EN;
> > writel(func_ctrl_reg_val, bus->base + ASPEED_I2C_FUN_CTRL_REG);
> > +
> > + bus->slave_state = ASPEED_I2C_SLAVE_INACTIVE;
> > }
> >
> > static int aspeed_i2c_reg_slave(struct i2c_client *client)
> > @@ -765,7 +767,6 @@ static int aspeed_i2c_reg_slave(struct i2c_client
> > *client)
> > __aspeed_i2c_reg_slave(bus, client->addr);
> >
> > bus->slave = client;
> > - bus->slave_state = ASPEED_I2C_SLAVE_INACTIVE;
> > spin_unlock_irqrestore(&bus->lock, flags);
> >
> > return 0;
> >
> >
>
> We tested both Jian's patch and Andrew's patch on our MCTP-i2c bus
> (ast2600 based BMC) and see both patches work well.
>
> We currently use upstream i2c-aspeed.c driver with the commit [1]
> backported. Without that commit, we frequently experienced the bus hang
> (due to bus arbitration) and it is unable to recover.
>
> But, by reverting that commit and with Jian or Andrew's patch, we see
> the bus could be able to recover so we think both changes are good.
>
> [1]
> https://github.com/AspeedTech-BMC/linux/commit/11a94e5918aa0f87c828d63fd254dd60ab2505e5
>
> Anyway, I would prefer Andrew's way because the bus->slave_state must
> always be reset to ASPEED_I2C_SLAVE_INACTIVE everytime
> __aspeed_i2c_reg_slave() is called.
Jian, what's your preference? Are you happy to do a v3 along the lines
of my suggestion above?
Otherwise Wolfram can take v2 and we can always do the cleanup in a
follow-up patch.
Andrew
> From: "Andrew Jeffery"<[email protected]>
> Date: Fri, Oct 6, 2023, 08:20
> Subject: [External] Re: [PATCH v2] i2c: aspeed: Fix i2c bus hang in slave read
> To: "Quan Nguyen"<[email protected]>, "Wolfram Sang"<[email protected]>, "Jian Zhang"<[email protected]>
> Cc: "Andi Shyti"<[email protected]>, "moderated list:ARM/ASPEED MACHINE SUPPORT"<[email protected]>, <[email protected]>, "moderated list:ARM/ASPEED I2C DRIVER"<[email protected]>, <[email protected]>, "open list"<[email protected]>, "Tommy Huang"<[email protected]>, "open list:ARM/ASPEED I2C DRIVER"<[email protected]>, <[email protected]>, <[email protected]>, <[email protected]>, "moderated list:ARM/ASPEED MACHINE SUPPORT"<[email protected]>, <[email protected]>
> On Thu, 2023-10-05 at 14:55 +0700, Quan Nguyen wrote:
> >
> > On 04/10/2023 13:08, Andrew Jeffery wrote:
> > > On Fri, 2023-09-29 at 09:39 +0200, Wolfram Sang wrote:
> > > > On Wed, Sep 27, 2023 at 11:42:43PM +0800, Jian Zhang wrote:
> > > > > When the `CONFIG_I2C_SLAVE` option is enabled and the device operates
> > > > > as a slave, a situation arises where the master sends a START signal
> > > > > without the accompanying STOP signal. This action results in a
> > > > > persistent I2C bus timeout. The core issue stems from the fact that
> > > > > the i2c controller remains in a slave read state without a timeout
> > > > > mechanism. As a consequence, the bus perpetually experiences timeouts.
> > > > >
> > > > > In this case, the i2c bus will be reset, but the slave_state reset is
> > > > > missing.
> > > > >
> > > > > Fixes: fee465150b45 ("i2c: aspeed: Reset the i2c controller when timeout occurs")
> > > > > Signed-off-by: Jian Zhang <[email protected]>
> > > >
> > > > Somebody wants to add tags here? I think it should go to my pull request
> > > > this week.
> > > >
> > >
> > > I've tested this patch applied on top of fee465150b45 on an AST2600 and
> > > the the system behaviour doesn't seem worse. However, I can still lock
> > > the bus up and trigger a hung task panic by surprise-unplugging things.
> > > I'll poke around to see if I can get to the bottom of that.
> > >
> > > Resetting the slave state makes sense, so with the above observation
> > > aside:
> > >
> > > Tested-by: Andrew Jeffery <[email protected]>
> > > Reviewed-by: Andrew Jeffery <[email protected]>
> > >
> > > That said I do wonder whether we should update the slave state in the
> > > same place we're updating the hardware state. It would cover off the
> > > gap identified by Jian if it were to ever occur anywhere else.
> > > Something like:
> > >
> > > diff --git a/drivers/i2c/busses/i2c-aspeed.c b/drivers/i2c/busses/i2c-
> > > aspeed.c
> > > index 5a416b39b818..28e2a5fc4528 100644
> > > --- a/drivers/i2c/busses/i2c-aspeed.c
> > > +++ b/drivers/i2c/busses/i2c-aspeed.c
> > > @@ -749,6 +749,8 @@ static void __aspeed_i2c_reg_slave(struct
> > > aspeed_i2c_bus *bus, u16 slave_addr)
> > > func_ctrl_reg_val = readl(bus->base + ASPEED_I2C_FUN_CTRL_REG);
> > > func_ctrl_reg_val |= ASPEED_I2CD_SLAVE_EN;
> > > writel(func_ctrl_reg_val, bus->base + ASPEED_I2C_FUN_CTRL_REG);
> > > +
> > > + bus->slave_state = ASPEED_I2C_SLAVE_INACTIVE;
> > > }
> > >
> > > static int aspeed_i2c_reg_slave(struct i2c_client *client)
> > > @@ -765,7 +767,6 @@ static int aspeed_i2c_reg_slave(struct i2c_client
> > > *client)
> > > __aspeed_i2c_reg_slave(bus, client->addr);
> > >
> > > bus->slave = client;
> > > - bus->slave_state = ASPEED_I2C_SLAVE_INACTIVE;
> > > spin_unlock_irqrestore(&bus->lock, flags);
> > >
> > > return 0;
> > >
> > >
> >
> > We tested both Jian's patch and Andrew's patch on our MCTP-i2c bus
> > (ast2600 based BMC) and see both patches work well.
> >
> > We currently use upstream i2c-aspeed.c driver with the commit [1]
> > backported. Without that commit, we frequently experienced the bus hang
> > (due to bus arbitration) and it is unable to recover.
> >
> > But, by reverting that commit and with Jian or Andrew's patch, we see
> > the bus could be able to recover so we think both changes are good.
> >
> > [1]
> > https://github.com/AspeedTech-BMC/linux/commit/11a94e5918aa0f87c828d63fd254dd60ab2505e5
> >
> > Anyway, I would prefer Andrew's way because the bus->slave_state must
> > always be reset to ASPEED_I2C_SLAVE_INACTIVE everytime
> > __aspeed_i2c_reg_slave() is called.
>
> Jian, what's your preference? Are you happy to do a v3 along the lines
> of my suggestion above?
Thanks, LGTM, I will send the patch v3.
Jian.
>
> Otherwise Wolfram can take v2 and we can always do the cleanup in a
> follow-up patch.
>
> Andrew