When RAM is slow (no training), it's possible to have the stack failing to negotiate configuration successfully.
The race condition is caused by not flushing cache before sending a configuration packet reply.
Although, cache is write-through, this needs to happen.
On XUSB do not allow multiple requests for CBW to be done.
This avoids an issue with some XHCI controllers and OS combos (e.g. ASMedia and Linux/Mac OS) which confuse that and concatenate an old CBW request with another write request (SCSI Write) and create a babble error (transmit overflow).
Some OSes do not adhere to limits reported by UMS gadget to them.
In such cases, make sure that user is notified about their host skipping the checks and sending "illegal" commands.
- Immediately exit if offset exceeds num of sectors.
- Correct cmnd type and checks.
- Inform user if ejected unsafely (while medium removal prevention is enabled).
The default interrupt moderation on XUSB controller was causing 4.62ms latency, hurting performance tremendously, especially in smaller usb packets (which are the norm).
This change brings it to parity with USB2 controller.
- Rework all timeouts to be more relaxed when doing big data transfers.
- Fix a bug where async transfer would timeout sooner instead of infinite tries.
Both showed up in Arch Linux, because of it's huge latency USB stack latency that can reach 1-2s.
The rework will let every OS work without adding additional wait time in the gadget loops.
BDK will allow developers to use the full collection of drivers,
with limited editing, if any, for making payloads for Nintendo Switch.
Using a single source for everything will also help decoupling
Switch specific code and easily port it to other Tegra X1/X1+ platforms.
And maybe even to lower targets.
Everything is now centrilized into bdk folder.
Every module or project can utilize it by simply including it.
This is just the start and it will continue to improve.
