On Tue, Jul 21, 2020 at 02:20:27PM +0200, SZIGETVÁRI János wrote: > Hi Alan, > > Alan Stern <stern@xxxxxxxxxxxxxxxxxxx> ezt írta (időpont: 2020. júl. > 20., H, 19:29): > > 4.19 is a very old kernel. You'd be a lot better off testing under a > > recent kernel, like 5.7. It also would be more convenient to do your > > development on a regular PC instead of a Raspberry Pi. You can use > > dummy-hcd for testing gadget drivers. > > Okay, I put together a Slackware-current based VM to conduct the > development work on. > I compiled a patched version of kernel 5.7.9 with all the necessary > config parameters to have USB gadget support with debugging enabled, > and also the dummy HCD. > > At first loaded the g_mass_storage module with a smaller ~700 MB iso > file, and it loaded seemingly fine, and functionality-wise it also > seemed to work fine, although there were frequent periodical debug > messages about the gadget returning command-failure status: > https://gist.github.com/jszigetvari/5bc4cdc8c55588907b71832558cb00a7 There failures are normal. They arise because the computer sends commands that the gadget driver doesn't understand or support, such as GET EVENT STATUS NOTIFICATION. > After that as a test I tried to dd all the data from /dev/sr1 (that's > the device file the gadget was assigned) to /dev/null. > The process completed without any user-visible errors, although the > failure status related messages kept on appearing: > https://gist.github.com/jszigetvari/ffa19850c94f12510c0b78ed5aaba2e1 > > Finally I tried to load the large, 11 GB CentOS iso file as the second > phase of the test. On the Raspberry Pi (with kernel 4.19) this was > when the crash happened, however this time it did not. > Originally Tiziano Bacocco's patch came around to be the time of the > 3.x kernels. The patch needed minimal adjustments to apply to 4.19 and > 5.4 and also to 5.7.9 (though there was a one line offset in > f_mass_storage.c). > https://gist.github.com/jszigetvari/5a9796c8af8a01c0edba6a8696540029 > > Although I have to admit that on the Pi Zero I was using the g_multi > kernel module, and not g_mass_storage I assumed they share the same > code base, so they should work the same way. > On the Pi zero I need the extra serial and Ethernet interface to have > access to the Pi Zero, as it doesn't have a wireless NIC, and that's > why I chose g_multi.ko. > > In the end I did a dd with the 11 GB iso, and to my surprise it worked > well. dd returned without any errors and it did show the true 11 GB > size when it finished. > In the kernel logs most log entries were about block reads, but the > failure status messages were still around. If there was no activity on > the device, then they accounted for the majority of the logs > generated: > https://gist.github.com/jszigetvari/a9225fa3173e6c7b7b50673a494d9c37 > > I have to admit, I stand puzzled why it worked on an x86_64-based > machine and crashed on the Pi Zero. The crash message on the Pi Zero indicated that there was a problem in the host controller driver (dwc2-hsotg), not in the mass-storage driver. > I guess the next step will be either: > * to try using g_multi instead of g_mass_storage on the test machine > * upgrade the Pi Zero to run kernel 5.4.x (AFAIK that's the latest one > can get from the Raspberry Pi Foundation's git repo), and recompile > that kernel version with the patch applied. (This will likely take a > few days.) Yes, that might indeed help. Alan Stern
