On Sat, Apr 27, 2019 at 12:45:00PM +0200, Greg Kroah-Hartman wrote: > On Sat, Apr 27, 2019 at 12:31:11PM +0300, Serge Semin wrote: > > On Sat, Apr 27, 2019 at 09:53:23AM +0200, Greg Kroah-Hartman wrote: > > > On Sat, Apr 27, 2019 at 10:39:41AM +0300, Serge Semin wrote: > > > > On Sat, Apr 27, 2019 at 09:32:00AM +0200, Greg Kroah-Hartman wrote: > > > > > On Sat, Apr 27, 2019 at 10:10:27AM +0300, Serge Semin wrote: > > > > > > On Sat, Apr 27, 2019 at 09:00:42AM +0200, Greg Kroah-Hartman wrote: > > > > > > > On Fri, Apr 26, 2019 at 01:46:50PM +0300, Serge Semin wrote: > > > > > > > > SMBus slave configuration is activated by CFG_SEL[1:0]=0x1 pins > > > > > > > > state. This is the mode the hub is supposed to be to let this driver > > > > > > > > work correctly. But a race condition might happen right after reset > > > > > > > > is cleared due to CFG_SEL[0] pin being multiplexed with SMBus SCL > > > > > > > > function. In case if the reset pin is handled by a i2c GPIO expander, > > > > > > > > which is also placed at the same i2c-bus segment as the usb251x > > > > > > > > SMB-interface connected to, then the hub reset clearance might > > > > > > > > cause the CFG_SEL[0] being latched in unpredictable state. So > > > > > > > > sometimes the hub configuration mode might be 0x1 (as expected), > > > > > > > > but sometimes being 0x0, which doesn't imply to have the hub SMBus-slave > > > > > > > > interface activated and consequently causes this driver failure. > > > > > > > > > > > > > > > > In order to fix the problem we must make sure the GPIO-reset chip doesn't > > > > > > > > reside the same i2c-bus segment as the SMBus-interface of the hub. If > > > > > > > > it doesn't, we can safely block the segment for the time the reset is > > > > > > > > cleared to prevent anyone generating a traffic at the i2c-bus SCL lane > > > > > > > > connected to the CFG_SEL[0] pin. But if it does, nothing we can do, so > > > > > > > > just return an error. If we locked the i2c-bus segment and tried to > > > > > > > > communicate with the GPIO-expander, it would cause a deadlock. If we didn't > > > > > > > > lock the i2c-bus segment, it would randomly cause the CFG_SEL[0] bit flip. > > > > > > > > > > > > > > > > Signed-off-by: Serge Semin <fancer.lancer@xxxxxxxxx> > > > > > > > > Signed-off-by: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx> > > > > > > > > > > > > > > I signed off on this? Where? When? > > > > > > > > > > > > > > never add a s-o-b line that you did not create, that implies a legal > > > > > > > agreement. > > > > > > > > > > > > > > > > > > > Ah, shit. Sorry. I should have added Acked-by. That's what I was going to do, > > > > > > since you already added the first version of this into linux-next tree. But > > > > > > apparently copy-pasted and left as is... > > > > > > > > > > If I have already applied a patch, I can't apply it again (or a > > > > > different version.) You need to send a fix-up patch for the reported > > > > > issue, not a whole new one as I can not go back in time and rewrite > > > > > history. > > > > > > > > > > thanks, > > > > > > > > > > greg k-h > > > > > > > > Ok. Should I send it in reply to this patch or as a completely separate one? > > > > > > A separate one. > > > > Just did this. > > > > Two more questions. > > > > Is there any online service or ready-to-use configs-gen-script to build > > patchsets locally with various kernel configs before sending them for review? > > By using this I'd most likely prevent such unfortunate mistakes in future > > patches. > > 'make allmodconfig' works for me. Also, the 0-day bot sometimes will > test patches sent to the mailing list before they are applied, but that > doesn't always work. > > > I suppose the situation like this isn't unique, and there are other patches > > submitted with possible compilation problems. In this case seeing the linux-next > > will be eventually integrated into the linux-stable, then at the point of those > > patches the kernel source code will be build-broken corresponding config. > > In this case wouldn't it be better to have some rebasable branch with build > > robot service and fix the problematic patches in-place instead of > > apply an incrementing commits? This also would keep the source code cleaner > > at that particular commits. > > As this passed the first round of 0-day testing in my rebasable branch, > I assumed it was good enough. But later testing found that not to be > the case, which is pretty rare. And I can't have public trees that are > rebasable as people work off of them, so we can just live with fixing > this up. > > thanks, > > greg k-h Ok. Thanks for clarification. Next time I'll use the allmodconfig for a generic build-test and, as there aren't other way, try to anticipate a possible configs disabling issues. -Sergey
