[U-Boot] [RFC] Make i2c probe opt-outable?
Hey all,
I'd like to propose making 'i2c probe' be a command that is opt-out'able. In the Linux Kernel the notion of probing for devices was abandoned a while ago due to, in short, devices misbehaving when randomly poked at. Over in omap24xx_i2c land we changed our probe method a while ago from an attempted read to an attempted write as some i2c devices would NAK the read. But now with the am33xx SoM family we have a new issue which is that attempting to write to an address doesn't immediately issue a NAK so probe sees all addresses as valid and in turn leaves the bus upset. I've worked around this for now by making i2c_probe use the read method instead, only on am33xx (so most devices would be spotted, but the ones that caused the initial change would not show up). But a possibly better solution is to just make the i2c probe command not implemented for am33xx (as you don't have to run i2c probe to try and use your device).
Thoughts? Thanks!
On Thursday 17 May 2012 14:43:45 Tom Rini wrote:
I'd like to propose making 'i2c probe' be a command that is opt-out'able. In the Linux Kernel the notion of probing for devices was abandoned a while ago due to, in short, devices misbehaving when randomly poked at. Over in omap24xx_i2c land we changed our probe method a while ago from an attempted read to an attempted write as some i2c devices would NAK the read. But now with the am33xx SoM family we have a new issue which is that attempting to write to an address doesn't immediately issue a NAK so probe sees all addresses as valid and in turn leaves the bus upset. I've worked around this for now by making i2c_probe use the read method instead, only on am33xx (so most devices would be spotted, but the ones that caused the initial change would not show up). But a possibly better solution is to just make the i2c probe command not implemented for am33xx (as you don't have to run i2c probe to try and use your device).
i've always seen the "i2c probe" command as a debugging tool, not something that you need to do to make things work. so along those lines, isn't it already optional ? if you don't like it, don't run it :). -mike
On 05/17/2012 07:48 PM, Mike Frysinger wrote:
On Thursday 17 May 2012 14:43:45 Tom Rini wrote:
I'd like to propose making 'i2c probe' be a command that is opt-out'able. In the Linux Kernel the notion of probing for devices was abandoned a while ago due to, in short, devices misbehaving when randomly poked at. Over in omap24xx_i2c land we changed our probe method a while ago from an attempted read to an attempted write as some i2c devices would NAK the read. But now with the am33xx SoM family we have a new issue which is that attempting to write to an address doesn't immediately issue a NAK so probe sees all addresses as valid and in turn leaves the bus upset. I've worked around this for now by making i2c_probe use the read method instead, only on am33xx (so most devices would be spotted, but the ones that caused the initial change would not show up). But a possibly better solution is to just make the i2c probe command not implemented for am33xx (as you don't have to run i2c probe to try and use your device).
i've always seen the "i2c probe" command as a debugging tool, not something that you need to do to make things work. so along those lines, isn't it already optional ? if you don't like it, don't run it :).
Including a command that doesn't work and saying "ah, just don't use that" is asking for trouble. I'm going down the "what changed in the IP block, really" rat-hole now (since I've got the original test working). But still, the kernel decided i2c probing is dangerous/unreliable, maybe we should follow, or at least allow boards to follow?
On Friday 18 May 2012 12:11:57 Tom Rini wrote:
On 05/17/2012 07:48 PM, Mike Frysinger wrote:
On Thursday 17 May 2012 14:43:45 Tom Rini wrote:
I'd like to propose making 'i2c probe' be a command that is opt-out'able. In the Linux Kernel the notion of probing for devices was abandoned a while ago due to, in short, devices misbehaving when randomly poked at. Over in omap24xx_i2c land we changed our probe method a while ago from an attempted read to an attempted write as some i2c devices would NAK the read. But now with the am33xx SoM family we have a new issue which is that attempting to write to an address doesn't immediately issue a NAK so probe sees all addresses as valid and in turn leaves the bus upset. I've worked around this for now by making i2c_probe use the read method instead, only on am33xx (so most devices would be spotted, but the ones that caused the initial change would not show up). But a possibly better solution is to just make the i2c probe command not implemented for am33xx (as you don't have to run i2c probe to try and use your device).
i've always seen the "i2c probe" command as a debugging tool, not something that you need to do to make things work. so along those lines, isn't it already optional ? if you don't like it, don't run it :).
Including a command that doesn't work and saying "ah, just don't use that" is asking for trouble. I'm going down the "what changed in the IP block, really" rat-hole now (since I've got the original test working).
there's plenty of commands in u-boot which are dangerous and should not be run lightly. i'm not sure we should special case this.
But still, the kernel decided i2c probing is dangerous/unreliable, maybe we should follow, or at least allow boards to follow?
i don't think "dangerous" is quite right, but certainly it's unreliable because i2c clients are dirt cheap and they aren't required by the protocol to be terribly sane. plenty of device drivers do probing once they've been told to connect to a specific address, but that's been gated somewhat.
also, the counter point is that the kernel still *allows* you to probe the entire i2c bus regardless of it being dangerous. the i2c-tools package is awesome for doing this sort of thing: probing an entire bus, attempting to dump register addresses of slaves en mass, etc... http://www.lm-sensors.org/wiki/I2CTools personally, i've used the "i2c probe" command before and found it very useful. -mike
Dear Tom Rini,
In message 4FB674CD.8030502@ti.com you wrote:
But still, the kernel decided i2c probing is dangerous/unreliable, maybe we should follow, or at least allow boards to follow?
The kernel is in a very different situation - it is supposed to run on perfect hardware.
In U-Boot, we have to dead with green hardware that is in a completely unknown state, or with hardware that _might_ be broken in one way or another. For us tools doing low-level accesses are invaluable - even if they include the risk to hang a system.
Best regards,
Wolfgang Denk
On 05/18/2012 11:26 AM, Wolfgang Denk wrote:
Dear Tom Rini,
In message4FB674CD.8030502@ti.com you wrote:
But still, the kernel decided i2c probing is dangerous/unreliable, maybe we should follow, or at least allow boards to follow?
The kernel is in a very different situation - it is supposed to run on perfect hardware.
In U-Boot, we have to dead with green hardware that is in a completely unknown state, or with hardware that _might_ be broken in one way or another. For us tools doing low-level accesses are invaluable - even if they include the risk to hang a system.
Right. I'm not suggesting removing i2c probe, I'm just suggesting making it opt-out'able. For example, on the beaglebone I can make i2c probe work as expected. It's just requring that we toggle the I2C_CON enable bit for some reason in probe, but not when we are actually trying to write to a bad address. I'd like to just #undef CONFIG_CMD_I2C_PROBE for the board and be done with it. Otherwise it lists (nearly) every address which is useless.
Dear Tom Rini,
In message 4FB69847.2090703@ti.com you wrote:
Right. I'm not suggesting removing i2c probe, I'm just suggesting making it opt-out'able. For example, on the beaglebone I can make i2c probe work as expected. It's just requring that we toggle the I2C_CON enable bit for some reason in probe, but not when we are actually trying to write to a bad address. I'd like to just #undef CONFIG_CMD_I2C_PROBE for the board and be done with it. Otherwise it lists (nearly) every address which is useless.
Can you use bit-banged I2C on these boards?
Best regards,
Wolfgang Denk
On Fri, May 18, 2012 at 11:52:06PM +0200, Wolfgang Denk wrote:
Dear Tom Rini,
In message 4FB69847.2090703@ti.com you wrote:
Right. I'm not suggesting removing i2c probe, I'm just suggesting making it opt-out'able. For example, on the beaglebone I can make i2c probe work as expected. It's just requring that we toggle the I2C_CON enable bit for some reason in probe, but not when we are actually trying to write to a bad address. I'd like to just #undef CONFIG_CMD_I2C_PROBE for the board and be done with it. Otherwise it lists (nearly) every address which is useless.
Can you use bit-banged I2C on these boards?
In this particular case it worked out to being that a previous change to the probe code violated the TRM, but didn't cause (obvious?) harm on older versions of the IP block but on newer ones (am33xx, omap4+) it does. I'll drop the idea of making i2c probe command opt-out.
participants (3)
-
Mike Frysinger -
Tom Rini -
Wolfgang Denk