On Mon, 1 May 2006, Pete Zaitcev wrote: > On Mon, 24 Apr 2006 11:55:51 -0400 (EDT), Alan Stern <[EMAIL PROTECTED]> > wrote: > > > Before submitting the patch, it seemed like a good idea to check with you > > and make sure it doesn't cause any problems. I'd appreciate it if you > > could try it out and let me know what happens. The patch is based on > > 2.6.17-rc2. > > Everything works fine here. I verified briefly that USBR do not get stuck > and that performace is adquate.
Good; thanks for testing. > The code looks fine though this concerns me a little: > > > + __le32 old_status, new_status, hw_status; > > + > > + /* Do this carefully because the hardware might > > + * be updating td->status right now. There's still > > + * a race because the controller might already have > > + * fetched the old value and be almost ready to > > + * overwrite the IOC with a new value. Nothing > > + * much we can do if that happens. */ > > + old_status = cpu_to_le32(status); > > + new_status = cpu_to_le32(status | TD_CTRL_IOC); > > + hw_status = cmpxchg(&qh->next_td->status, > > + old_status, new_status); > > + if (hw_status != old_status) > > + goto restart; > > + } It concerned me also. However it seemed important, and you left it out of your initial implementation. > The cmpxchg() is a second class citizen in platform API, so it may not > be implemented, or is implemented without any atomicity guarantees. > I'm pretty sure some CPUs cannot make these guarantees when they contend > with device DMA rather than other CPUs. > > I don't know what to do about this. It's tempting to ignore the issue > and say that owners of obscure platforms deserve what they get... My hope was that people wouldn't use UHCI controllers on those troublesome platforms. :-) More seriously, it is true that UHCI is confined to a relatively small set of platforms. Is there some way to discover at buildtime or runtime whether cmpxchg is atomic with respect to DMA access? (Should I post this question on LKML?) If it turns out not to be atomic, that code could simply be skipped. It would slow things down a little, since FSBR wouldn't get turned back on right away, but that's not as bad as overwriting a completion status. Better suboptimal than wrong. Alan Stern ------------------------------------------------------- Using Tomcat but need to do more? Need to support web services, security? Get stuff done quickly with pre-integrated technology to make your job easier Download IBM WebSphere Application Server v.1.0.1 based on Apache Geronimo http://sel.as-us.falkag.net/sel?cmd=lnk&kid=120709&bid=263057&dat=121642 _______________________________________________ linux-usb-devel@lists.sourceforge.net To unsubscribe, use the last form field at: https://lists.sourceforge.net/lists/listinfo/linux-usb-devel
