We have found a regression bug, where more than 512 URBs cannot be reliably submitted to XHCI. URBs beyond that return 0x00 instead of valid data in the buffer.
Our software works reliably on kernel versions through 6.4.x and fails on versions 6.5, 6.6, 6.7, and 6.8.0-rc6. This was discovered when Ubuntu recently updated their latest kernel package to version 6.5.
The issue is limited to the XHCI driver and appears to be isolated to this specific commit:
[ https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/dr... | https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/dr... ]
Attached is a test program that demonstrates the problem. We used a few different USB-to-Serial adapters with no driver installed as a convenient way to reproduce. We check the TRB debug information before and after to verify the actual number of allocated TRBs.
With some adapters on unaffected kernels, the TRB map gets expanded correctly. This directly corresponds to correct functional behavior. On affected kernels, the TRB ring does not expand, and our functional tests also will fail.
We don't know exactly why this happens. Some adapters do work correctly, so there seems to also be some subtle problem that was being masked by the liberal expansion of the TRB ring in older kernels. We also saw on one system that the TRB expansion did work correctly with one particular adapter. However, on all systems at least two adapters did exhibit the problem and fail.
Would it be possible to resolve this regression for the 6.8 release and backport the fix to versions 6.5, 6.6, and 6.7?
#regzbot ^introduced: f5af638f0609af889f15c700c60b93c06cc76675