This maybe? https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/scsi/scsi_lib.c?h=v5.10-rc6#n1816 UAS: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/usb/storage/uas.c?h=v5.10-rc6#n918 BOT (AFAICT): https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/scsi/hosts.c?h=v5.10-rc6#n466 It would explain why the issue is only triggered with UAS drives. The questions (from me) are: 1. From the scsi layer POV (as per what __scsi_init_queue() does), what/which should we use as dma_dev? 2. Do we really need to set dma_boundary in the UAS host template (to PAGE_SIZE - 1)? 3. Kind of the same question as #1: when we clamp hw_max_sectors to dma max mapping size, should the size actually be "the smaller one among dev and sysdev"? Or is one of the two sizes *always* the smaller one? On Tue, 1 Dec 2020 at 02:19, Hans de Goede <hdegoede@xxxxxxxxxx> wrote: > > Hi, > > On 11/30/20 6:20 PM, Alan Stern wrote: > > On Mon, Nov 30, 2020 at 02:36:38PM +0100, Hans de Goede wrote: > >> Hi, > >> > >> On 11/30/20 2:30 PM, Greg KH wrote: > >>> On Mon, Nov 30, 2020 at 02:23:48PM +0100, Hans de Goede wrote: > >>>> Hi, > >>>> > >>>> On 11/30/20 1:58 PM, Tom Yan wrote: > >>>>> It's merely a moving of comment moving for/and a no-behavioral-change > >>>>> adaptation for the reversion.> > >>>> > >>>> IMHO the revert of the troublesome commit and the other/new changes really > >>>> should be 2 separate commits. But I will let Alan and Greg have the final > >>>> verdict on this. > >>> > >>> I would prefer to just revert the commits and not do anything > >>> different/special here so late in the release cycle. > >>> > >>> So, if Alan agrees, I'll be glad to do them on my end, I just need the > >>> commit ids for them. > >> > >> The troublesome commit are (in reverse, so revert, order): > >> > >> 5df7ef7d32fe ("uas: bump hw_max_sectors to 2048 blocks for SS or faster drives") > >> 558033c2828f ("uas: fix sdev->host->dma_dev") > >> 0154012f8018 ("usb-storage: fix sdev->host->dma_dev") > >> > >> Alan, the reason for reverting these is that using scsi_add_host_with_dma() as the > >> last 2 patches do, with the dmadev argument of that call pointing to the device > >> for the XHCI controller is causing changes to the DMA settings of the XHCI controller > >> itself which is causing regressions in 5.10, see this email thread: > >> > >> https://lore.kernel.org/linux-usb/fde7e11f-5dfc-8348-c134-a21cb1116285@xxxxxxxxxx/T/#t > > > > It's hard to go wrong with reverting, so it's okay with me. > > > > Still, Hans, have you checked out the difference between the > > scsi_add_host() and scsi_add_host_with_dma() calls? It's just a matter > > of using dev vs. sysdev. In particular, have you checked to see what > > those two devices are on your system? > > Its not just dev vs sysdev, its iface->dev vs bus->sysdev, and I assume > that the latter is actually the XHCI controller. > > my vote goes to reverting to avoid the regression for 5.10, esp. since > this is a clean revert of 3 patches with nothing depending / building > on top of the reverted commits. > > Then for 5.11 we can retry to introduce similar changes. I would be happy > to try a new patch-set for 5.11. > > > It seems likely that if one of those calls messes up some DMA settings, > > the other one does too -- just maybe not settings that matter much. > > I'm not very familiar with all the DMA mapping / mask code, but AFAIK making > changes to the DMA settings of a child will not influence the parent. > > Where as when passing bus->sysdev, then changes are made to a device > which is shared with other devices on the bus, which is why we see > a regression in an USB NIC driver being triggered by the UAS driver > binding to a device (on the same bus). > > At least that is my interpretation of this. I bisected the regression > and that pointed at the UAS DMA change and reverting it fixes things, > confirming that I did not make any mistakes during the bisect. > > Regards, > > Hans >
