On Tue, Nov 5, 2013 at 6:06 AM, Jiri Kosina <jkosina@xxxxxxx> wrote: > On Thu, 31 Oct 2013, Bjorn Helgaas wrote: > >> > Attached is dmesg output leading to timeouts (that are cured by my >> > original patch in this thread) and lspci. >> >> I opened https://bugzilla.kernel.org/show_bug.cgi?id=64141 for this >> issue and attached your dmesg log and lspci output. >> >> > Please let me know if there is anything else I could do, or if you are >> > going to proceed with my patch adding the quirk. >> >> Your quirk keeps us from disabling MSIs on the device during >> enumeration. But even if the BIOS left MSIs enabled, there's nothing >> to field the MSI until after the driver claims the device. So I don't >> believe this has to be done as a quirk. It should work just as well >> to do something in the driver when it claims the device. >> >> I guess another way to say this is that I don't think we understand >> what the real problem is, and if we just add a quirk to work around >> it, we might miss the chance to fix the real problem, and we may never >> be able to remove the special-case code we're adding in the generic >> path. >> >> I know you said you tried doing something in the driver, and it didn't >> work. I don't know exactly what you tried, but twa_probe() looks >> strange to me. The other drivers I looked at do all their PCI >> initialization before the scsi_host_alloc() / scsi_add_host() / >> scsi_scan_host() stuff. But twa_probe() has PCI stuff scattered >> around between those three SCSI calls. In particular, it does the MSI >> setup way down near the end, after scsi_add_host(), which seems like >> just the sort of thing that could explain this problem. > > What I tried was patch below, but it didn't have any observable effect -- > the commands sent to the controller would still time out the same way. > > Debugging this is not really straightforward for me unfortunately, as I > don't own the system myself. > > I agree that we don't fully understand what is happening, but the quirk > was the only way I have been able to come up with to make the device > functioning again (apart from reverting d5dea7d95). > > Any other ideas are welcome. This patch looks like a good start, but there's a whole lot of other PCI-related initialization that I would suggest moving as well -- pci_request_regions(), ioremap(), pci_set_drvdata(), pci_enable_msi, request_irq(), etc. I would do this myself, but there are some pieces that don't look completely trivial, e.g., things like TW_DISABLE_INTERRUPTS() and twa_reset_sequence() don't look like they're SCSI-specific, but they are currently implemented using tw_dev, which looks like it's allocated by scsi_host_alloc(). Untangling all this looks like more work than I want to sign up to. But I really don't want to put the quirk in because it's just a quick hack that apparently just covers up bugs in the driver, and it will be an annoyance in the PCI core forever. Bjorn > --- > drivers/scsi/3w-9xxx.c | 10 +++++----- > 1 files changed, 5 insertions(+), 5 deletions(-) > > diff --git a/drivers/scsi/3w-9xxx.c b/drivers/scsi/3w-9xxx.c > index ba754c3..bad7faf 100644 > --- a/drivers/scsi/3w-9xxx.c > +++ b/drivers/scsi/3w-9xxx.c > @@ -2055,6 +2055,11 @@ static int __devinit twa_probe(struct pci_dev *pdev, const struct pci_device_id > goto out_disable_device; > } > > + /* Try to enable MSI */ > + if (use_msi && (pdev->device != PCI_DEVICE_ID_3WARE_9000) && > + !pci_enable_msi(pdev)) > + set_bit(TW_USING_MSI, &tw_dev->flags); > + > host = scsi_host_alloc(&driver_template, sizeof(TW_Device_Extension)); > if (!host) { > TW_PRINTK(host, TW_DRIVER, 0x24, "Failed to allocate memory for device extension"); > @@ -2134,11 +2139,6 @@ static int __devinit twa_probe(struct pci_dev *pdev, const struct pci_device_id > le32_to_cpu(*(int *)twa_get_param(tw_dev, 2, TW_INFORMATION_TABLE, > TW_PARAM_PORTCOUNT, TW_PARAM_PORTCOUNT_LENGTH))); > > - /* Try to enable MSI */ > - if (use_msi && (pdev->device != PCI_DEVICE_ID_3WARE_9000) && > - !pci_enable_msi(pdev)) > - set_bit(TW_USING_MSI, &tw_dev->flags); > - > /* Now setup the interrupt handler */ > retval = request_irq(pdev->irq, twa_interrupt, IRQF_SHARED, "3w-9xxx", tw_dev); > if (retval) { > > > -- > Jiri Kosina > SUSE Labs -- To unsubscribe from this list: send the line "unsubscribe linux-scsi" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
