On Tue, 11 Jun 2019, Christoph Hellwig wrote: > Hi Alan, > > thanks for the explanation. It seems like what usb wants is to: > > - set sg_tablesize to 1 for devices that can't handle scatterlist at all Hmmm. usb-storage (and possible other drivers too) currently handles such controllers by setting up an SG transfer as a series of separate URBs, one for each scatterlist entry. But this is not the same thing, for two reasons: It has less I/O overhead than setting sg_tablesize to 1 because it sets up the whole transfer as a single SCSI command, which requires much less time and traffic on the USB bus than sending multiple commands. It has that requirement about each scatterlist element except the last being a multiple of the maximum packet size in length. (This is because the USB protocol says that a transfer ends whenever a less-than-maximum-size packet is encountered.) We would like to avoid the extra I/O overhead for host controllers that can't handle SG. In fact, switching to sg_tablesize = 1 would probably be considered a regression. > - set the virt boundary as-is for devices supporting "basic" scatterlist, > although that still assumes they can rejiggle them because for example > you could still get a smaller than expected first segment ala (assuming > a 1024 byte packet size and thus 1023 virt_boundary_mask): > > | 0 .. 511 | 512 .. 1023 | 1024 .. 1535 | > > as the virt_bondary does not guarantee that the first segment is > the same size as all the mid segments. But that is exactly the problem we need to solve. The issue which prompted the commit this thread is about arose in a situation where the block layer set up a scatterlist containing buffer sizes something like: 4096 4096 1536 1024 and the maximum packet size was 1024. The situation was a little unusual, because it involved vhci-hcd (a virtual HCD). This doesn't matter much in normal practice because: Block devices normally have a block size of 512 bytes or more. Smaller values are very uncommon. So scatterlist element sizes are always divisible by 512. xHCI is the only USB host controller type with a maximum packet size larger than 512, and xHCI hardware can do full scatter-gather so it doesn't care what the buffer sizes are. So another approach would be to fix vhci-hcd and then trust that the problem won't arise again, for the reasons above. We would be okay so long as nobody tried to use a USB-SCSI device with a block size of 256 bytes or less. > - do not set any limit on xhci > > But that just goes back to the original problem, and that is that with > swiotlb we are limited in the total dma mapping size, and recent block > layer changes in the way we handle the virt_boundary mean we now build > much larger requests by default. For SCSI ULDs to take that into > account I need to call dma_max_mapping_size() and use that as the > upper bound for the request size. My plan is to do that in scsi_lib.c, > but for that we need to expose the actual struct device that the dma > mapping is perfomed on to the scsi layer. If that device is different > from the sysfs hierchary struct device, which it is for usb the ULDD > needs to scsi_add_host_with_dma and pass the dma device as well. How > do I get at the dma device (aka the HCDs pci_dev or similar) from > usb-storage/uas? >From usb_stor_probe2(): us->pusb_dev->bus->sysdev. >From uas_probe(): udev->bus->sysdev. The ->sysdev field points to the device used for DMA mapping. It is often the same as ->controller, but sometimes it is ->controller->parent because of the peculiarities of some platforms. Alan Stern
