increase xhci bulk transfer sgl count
The xhci driver today defaults to using DMA attributes of up to 8 entries when performing bulk transfers. We had examples where the OS wanted 9 sgl entries for such a transfer. If for some reason the system cannot satisfy this number of transfers, the attempt to bind DMA memory will fail because the xhci driver does not support falling back to using DMA windows to perform the bulk transfer. As such, we should increase the default number of SGLs that we use. The upper bound on such a number (as described in the code) is based on the ring size. We use page size rings which means 4k on x86. When divided by the size of a ring entry, this gives us 256 entries in the ring. Thus we need to determine a number that is flexible enough, without going over the total size of the ring for a given number . For the time being we've chosen 64 as a middle ground.
Updated by Michal Nowak about 4 years ago
We recently had an user bug report regarding a similar issue when the user tried to access camera with gphoto2 app which is itself using libusb: https://openindiana.org/pipermail/openindiana-discuss/2018-May/022152.html.
Updated by Electric Monk over 3 years ago
- Status changed from New to Closed
commit 2aba3acda67326648fd60aaf2bfb4e18ee8c04ed Author: Robert Mustacchi <firstname.lastname@example.org> Date: 2018-12-01T05:49:03.000Z 9816 Multi-TRB xhci transfers should use event data 9817 xhci needs to always set slot context 8550 increase xhci bulk transfer sgl count 9818 xhci_transfer_get_tdsize can return values that are too large Reviewed by: Alex Wilson <email@example.com> Reviewed by: Jerry Jelinek <firstname.lastname@example.org> Approved by: Joshua M. Clulow <email@example.com>