Re: FYI: RPi4B's and such: VLI_SS_BULK_OUT_BUG quirk handling vs. VL805 (linux example)

On 2022-Sep-17, at 23:05, Hans Petter Selasky <hps@selasky.org> wrote:

> On 9/17/22 18:09, Mark Millard wrote:
>> https://github.com/raspberrypi/linux/pull/5173 reports:
>> QUOTE
>> After several months back-and-forth with VIA, we have a candidate root-cause for #4844 and a suggested fix. My known-bad pendrive now gets written to endlessly without suffering data corruption.
>> END QUOTE
>> and:
>> QUOTE
>> usb: xhci: expand mitigations for VLI_SS_BULK_OUT_BUG quirk … c9e051a
>> The VL805 can cause data corruption if a SS Bulk OUT endpoint enters a
>> flow-control condition and there are TRBs in the transfer ring that are
>> not an integral size of wMaxPacket and the endpoint is behind one or more
>> hubs.
>> This is frequently the case encountered when FAT32 filesystems are
>> present on mass-storage devices with cluster sizes of 1 sector, and the
>> filesystem is being written to with an aggregate of small files.
>> The initial implementation of this quirk separated TRBs that didn't
>> adhere to this limitation into two - the first a multiple of wMaxPacket
>> and the second the 512-byte remainder - in an attempt to force TD
>> fragments to align with packet boundaries. This reduced the incidence
>> rate of data corruption but did not resolve it.
>> The fix as recommended by VIA is to disable bursts if this sequence of
>> TRBs can occur.
>> Limit turning off bursts to just USB mass-storage devices by searching
>> the device's configuration for an interface with a class type of
>> USB_CLASS_MASS_STORAGE.
>> Signed-off-by: Jonathan Bell <jonathan@raspberrypi.com>
>> END QUOTE
>> and the drivers/usb/host/xhci-mem.c change has the comments:
>> 	/*
>> 	 * VL805 errata - Bulk OUT bursts to superspeed mass-storage
>> 	 * devices behind hub ports can cause data corruption with
>> 	 * non-wMaxPacket-multiple transfers.
>> 	 */
>> . . .
>> 		/*
>> 		 * Slight hack - look at interface altsetting 0, which
>> 		 * should be the UMS bulk-only interface. If the class
>> 		 * matches, then we disable out bursts for all OUT
>> 		 * endpoints because endpoint assignments may change
>> 		 * between alternate settings.
>> 		 */
>> ===
>> Mark Millard
>> marklmi at yahoo.com
> 
> Hi,
> 
> It doesn't make sense that this only applies to USB mass storage devices. Doesn't it apply to all SuperSpeed USB BULK transfers?

No clue.

> What about other OS'es, do they also get a fix?
> 

I doubt that the raspberry pi folks will go around fixing non-linux
kernels directly. I do see evidence of the XHCI_VLI_SS_BULK_OUT_BUG
name existing in some other linux kernels and past adjustments by
them to track past changes by the raspberry pi folks.

However,

https://elixir.bootlin.com/linux/latest/A/ident/XHCI_VLI_SS_BULK_OUT_BUG

does not find the symbol so linux does not have it in general.

(I'd not known of renaming or the like.)


===
Mark Millard
marklmi at yahoo.com