| Message ID | 5463A381.9060305@gmail.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On 11/12/2014 10:14 AM, Frank Rowand wrote: > On 11/10/2014 7:20 PM, Frank Rowand wrote: >> On 11/10/2014 6:07 PM, Stephen Boyd wrote: >>> On 11/10/2014 05:56 PM, Frank Rowand wrote: >>>> On 11/10/2014 11:42 AM, Stephen Boyd wrote: >>>>> diff --git a/drivers/tty/serial/msm_serial.c b/drivers/tty/serial/msm_serial.c >>>>> index 09364dd8cf3a..d1bc6b6cbc70 100644 >>>>> --- a/drivers/tty/serial/msm_serial.c >>>>> +++ b/drivers/tty/serial/msm_serial.c >>>>> @@ -1046,14 +1046,14 @@ static int msm_serial_probe(struct platform_device *pdev) >>>>> const struct of_device_id *id; >>>>> int irq, line; >>>>> >>>>> - if (pdev->id == -1) >>>>> - pdev->id = atomic_inc_return(&msm_uart_next_id) - 1; >>>>> - >>>>> if (pdev->dev.of_node) >>>>> line = of_alias_get_id(pdev->dev.of_node, "serial"); >>>>> else >>>>> line = pdev->id; >>>>> >>>>> + if (line < 0) >>>>> + line = atomic_inc_return(&msm_uart_next_id) - 1; >>>>> + >>>>> if (unlikely(line < 0 || line >= UART_NR)) >>>> Then this original check for "line < 0" can also be removed. >>>> >>>> >>> Well this matches what was there before. It would do atomic_inc_return >>> if the line was negative and then still check for a negative value. I >>> don't mind removing it though. Perhaps we should use an ida?: >>> >> OK, you are right. If (pdev->id < -1) and (!pdev->dev.of_node) then >> the check is still needed. >> >> You could use an ida. Some drivers use a bit map. I really don't think >> this should become a complicated algorithm though. If the rule is that >> either all UARTS have an alias, or no UART has an alias, then I think >> the patch could be something like the following. >> >> This combines your original patch, plus your fix patch, plus making >> the aliases all or nothing. Not tested, not even compiled. >> >> What do you think? > < snip - previous version of patch > > > Previous version of the patch did not protect against no alias, followed > by alias. > > > > From: Frank Rowand <frank.rowand@sonymobile.com> > > This patch is intended to show roughly what an implementation of an all or > nothing policy for using serialN aliases would look like. It is intended > simply to help determine whether this policy should be used. > > Combines Stephen's first patch, Stephen's fix patch (to make the serialN > alias optional), plus makes use of the serialN alias all or nothing (either > all ports have an alias or none do). > > This is not a correct implementation because the atomic_read(&msm_uart_next_id) > does not protect against concurrency; instead a lock should be used. > > Not tested, not even compiled. Sorry, I'm sort of lost. If there are serial aliases in the dts file, then we should alias all of the serial ports. If there aren't aliases then we're backwards compatible with the dts we have now and we'll do dynamic generation. Putting code into the driver to validate that this is true is not the job of the driver. If anything, it should validated when the dts file is created. If one day we screw up and have a dts file with such a bad configuration we'll have to work around it, but until that day comes I'd rather not think about it.
On 11/13/2014 11:31 AM, Stephen Boyd wrote: > On 11/12/2014 10:14 AM, Frank Rowand wrote: >> On 11/10/2014 7:20 PM, Frank Rowand wrote: >>> On 11/10/2014 6:07 PM, Stephen Boyd wrote: >>>> On 11/10/2014 05:56 PM, Frank Rowand wrote: >>>>> On 11/10/2014 11:42 AM, Stephen Boyd wrote: >>>>>> diff --git a/drivers/tty/serial/msm_serial.c b/drivers/tty/serial/msm_serial.c >>>>>> index 09364dd8cf3a..d1bc6b6cbc70 100644 >>>>>> --- a/drivers/tty/serial/msm_serial.c >>>>>> +++ b/drivers/tty/serial/msm_serial.c >>>>>> @@ -1046,14 +1046,14 @@ static int msm_serial_probe(struct platform_device *pdev) >>>>>> const struct of_device_id *id; >>>>>> int irq, line; >>>>>> >>>>>> - if (pdev->id == -1) >>>>>> - pdev->id = atomic_inc_return(&msm_uart_next_id) - 1; >>>>>> - >>>>>> if (pdev->dev.of_node) >>>>>> line = of_alias_get_id(pdev->dev.of_node, "serial"); >>>>>> else >>>>>> line = pdev->id; >>>>>> >>>>>> + if (line < 0) >>>>>> + line = atomic_inc_return(&msm_uart_next_id) - 1; >>>>>> + >>>>>> if (unlikely(line < 0 || line >= UART_NR)) >>>>> Then this original check for "line < 0" can also be removed. >>>>> >>>>> >>>> Well this matches what was there before. It would do atomic_inc_return >>>> if the line was negative and then still check for a negative value. I >>>> don't mind removing it though. Perhaps we should use an ida?: >>>> >>> OK, you are right. If (pdev->id < -1) and (!pdev->dev.of_node) then >>> the check is still needed. >>> >>> You could use an ida. Some drivers use a bit map. I really don't think >>> this should become a complicated algorithm though. If the rule is that >>> either all UARTS have an alias, or no UART has an alias, then I think >>> the patch could be something like the following. >>> >>> This combines your original patch, plus your fix patch, plus making >>> the aliases all or nothing. Not tested, not even compiled. >>> >>> What do you think? >> < snip - previous version of patch > >> >> Previous version of the patch did not protect against no alias, followed >> by alias. >> >> >> >> From: Frank Rowand <frank.rowand@sonymobile.com> >> >> This patch is intended to show roughly what an implementation of an all or >> nothing policy for using serialN aliases would look like. It is intended >> simply to help determine whether this policy should be used. >> >> Combines Stephen's first patch, Stephen's fix patch (to make the serialN >> alias optional), plus makes use of the serialN alias all or nothing (either >> all ports have an alias or none do). >> >> This is not a correct implementation because the atomic_read(&msm_uart_next_id) >> does not protect against concurrency; instead a lock should be used. >> >> Not tested, not even compiled. > > Sorry, I'm sort of lost. If there are serial aliases in the dts file, > then we should alias all of the serial ports. If there aren't aliases > then we're backwards compatible with the dts we have now and we'll do > dynamic generation. Putting code into the driver to validate that this > is true is not the job of the driver. If anything, it should validated > when the dts file is created. If one day we screw up and have a dts file > with such a bad configuration we'll have to work around it, but until > that day comes I'd rather not think about it. Maybe I did not understand when you said "Perhaps we should use an ida". That sentence led me to think the driver should check for misconfiguration. The case I was trying to handle was if there was at least one serialN alias and at least one UART without an alias. For example, if there are three UARTs (serial_a, serial_b, serial_c, probed in that order) and one alias (serial0 = &serial_c;) then the result would be: serial_a line 0 (from msm_uart_next_id) serial_b line 1 (from msm_uart_next_id) serial_c line 0 (from the alias) Two UARTs probed with line == 0. This is an error. Most of the serial drivers don't check for this type of bad configuration. Some drivers keep a bit map of which lines have been used. I'm not sure what they do in case of a conflict (I did not read to that level of detail). I thought you were suggesting the driver check for the bad configuration, so I was proposing a somewhat simple way of forcing a boot error for the bad configuration. Since you are not suggesting the driver check for the bad configuration, you can ignore my proposal. I agree that it is ok for the driver to expect the board dts to be correct. The problem should be detected by the dts author on first boot as part of normal bring up testing, and then corrected. -Frank
On 11/13/2014 04:46 PM, Frank Rowand wrote: > On 11/13/2014 11:31 AM, Stephen Boyd wrote: >> Sorry, I'm sort of lost. If there are serial aliases in the dts file, >> then we should alias all of the serial ports. If there aren't aliases >> then we're backwards compatible with the dts we have now and we'll do >> dynamic generation. Putting code into the driver to validate that >> this is true is not the job of the driver. If anything, it should >> validated when the dts file is created. If one day we screw up and >> have a dts file with such a bad configuration we'll have to work >> around it, but until that day comes I'd rather not think about it. > Maybe I did not understand when you said "Perhaps we should use an ida". > That sentence led me to think the driver should check for misconfiguration. > The case I was trying to handle was if there was at least one serialN > alias and at least one UART without an alias. For example, if there > are three UARTs (serial_a, serial_b, serial_c, probed in that order) > and one alias (serial0 = &serial_c;) then the result would be: > > serial_a line 0 (from msm_uart_next_id) > serial_b line 1 (from msm_uart_next_id) > serial_c line 0 (from the alias) > > Two UARTs probed with line == 0. This is an error. > > Most of the serial drivers don't check for this type of bad configuration. > Some drivers keep a bit map of which lines have been used. I'm not sure > what they do in case of a conflict (I did not read to that level of detail). > > I thought you were suggesting the driver check for the bad configuration, > so I was proposing a somewhat simple way of forcing a boot error for the > bad configuration. > > Since you are not suggesting the driver check for the bad configuration, > you can ignore my proposal. I agree that it is ok for the driver to > expect the board dts to be correct. The problem should be detected by > the dts author on first boot as part of normal bring up testing, and > then corrected. > Ah ok. I was just saying we could use an ida instead of an atomic increment so that this driver works properly with driver binding/unbinding, otherwise the line number keeps increasing and quickly goes beyond the static array of ports (which I still don't understand why we have at all btw).
Stephen Boyd <sboyd@codeaurora.org> writes: > On 11/13/2014 04:46 PM, Frank Rowand wrote: >> On 11/13/2014 11:31 AM, Stephen Boyd wrote: >>> Sorry, I'm sort of lost. If there are serial aliases in the dts file, >>> then we should alias all of the serial ports. If there aren't aliases >>> then we're backwards compatible with the dts we have now and we'll do >>> dynamic generation. Putting code into the driver to validate that >>> this is true is not the job of the driver. If anything, it should >>> validated when the dts file is created. If one day we screw up and >>> have a dts file with such a bad configuration we'll have to work >>> around it, but until that day comes I'd rather not think about it. >> Maybe I did not understand when you said "Perhaps we should use an ida". >> That sentence led me to think the driver should check for misconfiguration. >> The case I was trying to handle was if there was at least one serialN >> alias and at least one UART without an alias. For example, if there >> are three UARTs (serial_a, serial_b, serial_c, probed in that order) >> and one alias (serial0 = &serial_c;) then the result would be: >> >> serial_a line 0 (from msm_uart_next_id) >> serial_b line 1 (from msm_uart_next_id) >> serial_c line 0 (from the alias) >> >> Two UARTs probed with line == 0. This is an error. >> >> Most of the serial drivers don't check for this type of bad configuration. >> Some drivers keep a bit map of which lines have been used. I'm not sure >> what they do in case of a conflict (I did not read to that level of detail). >> >> I thought you were suggesting the driver check for the bad configuration, >> so I was proposing a somewhat simple way of forcing a boot error for the >> bad configuration. >> >> Since you are not suggesting the driver check for the bad configuration, >> you can ignore my proposal. I agree that it is ok for the driver to >> expect the board dts to be correct. The problem should be detected by >> the dts author on first boot as part of normal bring up testing, and >> then corrected. >> > > Ah ok. I was just saying we could use an ida instead of an atomic > increment so that this driver works properly with driver > binding/unbinding, otherwise the line number keeps increasing and > quickly goes beyond the static array of ports (which I still don't > understand why we have at all btw). Due to the length of the thread, I haven't followed all the details, and I suspect Greg hasn't either, so I'm not sure if you're discssuing what the right fix is for what's in -next (still broken[1], or what should be done with the device board files. If the fix from earlier in this thread is still the right one for fixing -next, could you repost it separately for Greg to queue/squash and for me to re-test (if needed.) Thanks, Kevin [1] http://lists.linaro.org/pipermail/kernel-build-reports/2014-November/006298.html
On 11/14, Kevin Hilman wrote: > > Due to the length of the thread, I haven't followed all the details, and > I suspect Greg hasn't either, so I'm not sure if you're discssuing what > the right fix is for what's in -next (still broken[1], or what should be done > with the device board files. > > If the fix from earlier in this thread is still the right one for fixing > -next, could you repost it separately for Greg to queue/squash and for > me to re-test (if needed.) > No problem. I'll pick up your tested-by and resend.
Index: linux/drivers/tty/serial/msm_serial.c =================================================================== --- linux.orig/drivers/tty/serial/msm_serial.c +++ linux/drivers/tty/serial/msm_serial.c @@ -1044,17 +1044,31 @@ static int msm_serial_probe(struct platf struct resource *resource; struct uart_port *port; const struct of_device_id *id; - int irq; + int irq, line; + static int no_prev_alias = 1; - if (pdev->id == -1) - pdev->id = atomic_inc_return(&msm_uart_next_id) - 1; + if (pdev->dev.of_node) { + line = of_alias_get_id(pdev->dev.of_node, "serial"); + if (line < 0 && no_prev_alias) { + line = atomic_inc_return(&msm_uart_next_id) - 1; + } else { + no_prev_alias = 0; + if (atomic_read(&msm_uart_next_id)) + line = -ENXIO; + } + } else { + if (pdev->id < 0 && no_prev_alias) + line = atomic_inc_return(&msm_uart_next_id) - 1; + else + line = pdev->id; + } - if (unlikely(pdev->id < 0 || pdev->id >= UART_NR)) + if (unlikely(line < 0 || line >= UART_NR)) return -ENXIO; - dev_info(&pdev->dev, "msm_serial: detected port #%d\n", pdev->id); + dev_info(&pdev->dev, "msm_serial: detected port #%d\n", line); - port = get_port_from_line(pdev->id); + port = get_port_from_line(line); port->dev = &pdev->dev; msm_port = UART_TO_MSM(port);