| Message ID | 1389265885-26777-1-git-send-email-mike.looijmans@topic.nl (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Thu, Jan 09, 2014 at 12:11:25PM +0100, Mike Looijmans wrote: > When a signal is caught while the i2c-davinci bus driver is transferring, > the drive just "abandons" the transfer and leaves the controller to fend > for itself. The next I2C transaction will find the controller in an > undefined state and often results in a stream of "initiating i2c bus recovery" > messages until the controller arrives in a defined state. This behaviour > also sends out "half" or possibly even mixed messages to I2C client > devices which may put them in an undesired state as well. > > This patch fixes this issue by always attempting to finish the current > transaction, and then check on a pending signal. It either reports > success if all data has been transferred, or it returns failure when > the transaction was aborted. This keeps the controller in a defined > state, and is also much friendlier towards client devices, because > it will only send complete messages. Even more, you should complete the whole transfer. There are devices where things can really go wrong if you send a half-complete command and then start with the next one. So, not checking signals at all is the way to go for I2C drivers. There is some cruft left, so I am happy about patches fixing that, with testing on real HW. Like yours here.
?On 03/09/2014 09:21 PM, Wolfram Sang wrote: > On Thu, Jan 09, 2014 at 12:11:25PM +0100, Mike Looijmans wrote: >> When a signal is caught while the i2c-davinci bus driver is transferring, >> the drive just "abandons" the transfer and leaves the controller to fend >> for itself. The next I2C transaction will find the controller in an >> undefined state and often results in a stream of "initiating i2c bus recovery" >> messages until the controller arrives in a defined state. This behaviour >> also sends out "half" or possibly even mixed messages to I2C client >> devices which may put them in an undesired state as well. >> >> This patch fixes this issue by always attempting to finish the current >> transaction, and then check on a pending signal. It either reports >> success if all data has been transferred, or it returns failure when >> the transaction was aborted. This keeps the controller in a defined >> state, and is also much friendlier towards client devices, because >> it will only send complete messages. > > Even more, you should complete the whole transfer. There are devices > where things can really go wrong if you send a half-complete command and > then start with the next one. So, not checking signals at all is the way > to go for I2C drivers. There is some cruft left, so I am happy about > patches fixing that, with testing on real HW. Like yours here. I agree. I know the Zynq (using a cadence controller) also lets signals interrupt I2C transfers, so I'll propose a patch to Xilinx and CC to you and linux-i2c to completely remove signal handling from that driver as well. Mike Met vriendelijke groet / kind regards, Mike Looijmans TOPIC Embedded Systems Eindhovenseweg 32-C, NL-5683 KH Best Postbus 440, NL-5680 AK Best Telefoon: (+31) (0) 499 33 69 79 Telefax: (+31) (0) 499 33 69 70 E-mail: mike.looijmans@topic.nl Website: www.topic.nl Please consider the environment before printing this e-mail
> >Even more, you should complete the whole transfer. There are devices > >where things can really go wrong if you send a half-complete command and > >then start with the next one. So, not checking signals at all is the way > >to go for I2C drivers. There is some cruft left, so I am happy about > >patches fixing that, with testing on real HW. Like yours here. > > I agree. > > I know the Zynq (using a cadence controller) also lets signals > interrupt I2C transfers, so I'll propose a patch to Xilinx and CC to > you and linux-i2c to completely remove signal handling from that > driver as well. Cool, thanks!
> > >Even more, you should complete the whole transfer. There are devices > > >where things can really go wrong if you send a half-complete command and > > >then start with the next one. So, not checking signals at all is the way > > >to go for I2C drivers. There is some cruft left, so I am happy about > > >patches fixing that, with testing on real HW. Like yours here. > > > > I agree. > > > > I know the Zynq (using a cadence controller) also lets signals > > interrupt I2C transfers, so I'll propose a patch to Xilinx and CC to > > you and linux-i2c to completely remove signal handling from that > > driver as well. > > Cool, thanks! Are you going to update the davinci patch as well?
?On 03/10/2014 04:24 PM, Wolfram Sang wrote: > >>>> Even more, you should complete the whole transfer. There are devices >>>> where things can really go wrong if you send a half-complete command and >>>> then start with the next one. So, not checking signals at all is the way >>>> to go for I2C drivers. There is some cruft left, so I am happy about >>>> patches fixing that, with testing on real HW. Like yours here. >>> >>> I agree. >>> >>> I know the Zynq (using a cadence controller) also lets signals >>> interrupt I2C transfers, so I'll propose a patch to Xilinx and CC to >>> you and linux-i2c to completely remove signal handling from that >>> driver as well. >> >> Cool, thanks! > > Are you going to update the davinci patch as well? An amended patch is on its way now. I forgot to set the subject to "PATCHv2" though. Mike. Met vriendelijke groet / kind regards, Mike Looijmans TOPIC Embedded Systems Eindhovenseweg 32-C, NL-5683 KH Best Postbus 440, NL-5680 AK Best Telefoon: (+31) (0) 499 33 69 79 Telefax: (+31) (0) 499 33 69 70 E-mail: mike.looijmans@topic.nl Website: www.topic.nl Please consider the environment before printing this e-mail
diff --git a/drivers/i2c/busses/i2c-davinci.c b/drivers/i2c/busses/i2c-davinci.c index 132369f..102673b 100644 --- a/drivers/i2c/busses/i2c-davinci.c +++ b/drivers/i2c/busses/i2c-davinci.c @@ -372,9 +372,9 @@ i2c_davinci_xfer_msg(struct i2c_adapter *adap, struct i2c_msg *msg, int stop) flag |= DAVINCI_I2C_MDR_STP; davinci_i2c_write_reg(dev, DAVINCI_I2C_MDR_REG, flag); - r = wait_for_completion_interruptible_timeout(&dev->cmd_complete, + r = wait_for_completion_timeout(&dev->cmd_complete, dev->adapter.timeout); - if (r == 0) { + if (unlikely(r == 0)) { dev_err(dev->dev, "controller timed out\n"); davinci_i2c_recover_bus(dev); i2c_davinci_init(dev); @@ -384,7 +384,6 @@ i2c_davinci_xfer_msg(struct i2c_adapter *adap, struct i2c_msg *msg, int stop) if (dev->buf_len) { /* This should be 0 if all bytes were transferred * or dev->cmd_err denotes an error. - * A signal may have aborted the transfer. */ if (r >= 0) { dev_err(dev->dev, "abnormal termination buf_len=%i\n", @@ -436,22 +435,30 @@ i2c_davinci_xfer(struct i2c_adapter *adap, struct i2c_msg msgs[], int num) ret = i2c_davinci_wait_bus_not_busy(dev, 1); if (ret < 0) { dev_warn(dev->dev, "timeout waiting for bus ready\n"); - return ret; + goto error; } for (i = 0; i < num; i++) { + if (signal_pending(current)) { + dev_dbg(dev->dev, "%s [%d/%d] %#x ERESTARTSYS\n", + __func__, i + 1, num, msgs[i].addr); + ret = -ERESTARTSYS; + goto error; + } ret = i2c_davinci_xfer_msg(adap, &msgs[i], (i == (num - 1))); - dev_dbg(dev->dev, "%s [%d/%d] ret: %d\n", __func__, i + 1, num, - ret); + dev_dbg(dev->dev, "%s [%d/%d] %#x ret: %d\n", __func__, i + 1, + num, msgs[i].addr, ret); if (ret < 0) - return ret; + goto error; } + ret = num; +error: #ifdef CONFIG_CPU_FREQ complete(&dev->xfr_complete); #endif - return num; + return ret; } static u32 i2c_davinci_func(struct i2c_adapter *adap)
When a signal is caught while the i2c-davinci bus driver is transferring, the drive just "abandons" the transfer and leaves the controller to fend for itself. The next I2C transaction will find the controller in an undefined state and often results in a stream of "initiating i2c bus recovery" messages until the controller arrives in a defined state. This behaviour also sends out "half" or possibly even mixed messages to I2C client devices which may put them in an undesired state as well. This patch fixes this issue by always attempting to finish the current transaction, and then check on a pending signal. It either reports success if all data has been transferred, or it returns failure when the transaction was aborted. This keeps the controller in a defined state, and is also much friendlier towards client devices, because it will only send complete messages. Before this patch, reading an I2C device in a loop and interrupting it often resulted in a "initiating i2c bus recovery" storm and not being able to communicate via I2C for several seconds. With this patch, the userspace call simply returns EINTR and the next I2C transaction succeeds without errors. Signed-off-by: Mike Looijmans <mike.looijmans@topic.nl> --- drivers/i2c/busses/i2c-davinci.c | 23 +++++++++++++++-------- 1 file changed, 15 insertions(+), 8 deletions(-)