Message-ID: <20220318210947.GA845994@bhelgaas>
Date:   Fri, 18 Mar 2022 16:09:47 -0500
From:   Bjorn Helgaas <helgaas@...nel.org>
To:     Andy Shevchenko <andy.shevchenko@...il.com>
Cc:     Bjorn Helgaas <bhelgaas@...gle.com>,
        linux-pci <linux-pci@...r.kernel.org>,
        Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
        micklorain@...tonmail.com,
        Alex Williamson <alex.williamson@...hat.com>
Subject: Re: [PATCH v1 1/1] PCI: Enable INTx quirk for ATI PCIe-USB adapter

On Fri, Mar 18, 2022 at 08:06:15PM +0200, Andy Shevchenko wrote:
> On Fri, Mar 18, 2022 at 11:47:40AM -0500, Bjorn Helgaas wrote:
> > On Fri, Mar 18, 2022 at 12:42:18PM +0200, Andy Shevchenko wrote:
> > > On Thu, Mar 17, 2022 at 11:12 PM Bjorn Helgaas <helgaas@...nel.org> wrote:
> > > > On Thu, Mar 17, 2022 at 10:59:28AM +0200, Andy Shevchenko wrote:
> > > > > On Wed, Mar 16, 2022 at 04:15:48PM -0500, Bjorn Helgaas wrote:
> > > > > > On Wed, Mar 16, 2022 at 06:12:19PM +0200, Andy Shevchenko wrote:
> > > > > > > On Wed, Mar 16, 2022 at 06:52:09AM -0500, Bjorn Helgaas wrote:
> > > > > > > > On Wed, Mar 16, 2022 at 12:27:57PM +0200, Andy Shevchenko wrote:
> > > > > > > > > On Tue, Mar 15, 2022 at 03:22:31PM -0500, Bjorn Helgaas wrote:
> > > > > > > > > > On Tue, Mar 15, 2022 at 12:09:08PM +0200, Andy Shevchenko wrote:
> > > > > > > > > > > On Mon, Mar 14, 2022 at 02:42:53PM -0500, Bjorn Helgaas wrote:
> > > > > > > > > > > > On Mon, Mar 14, 2022 at 12:14:48PM +0200, Andy Shevchenko wrote:
> > > > > > > > > > > > > ATI PCIe-USB adapter advertises MSI, but it doesn't work
> > > > > > > > > > > > > if INTx is disabled.  Enable the respective quirk as
> > > > > > > > > > > > > it's done for other ATI devices on this chipset,
> > > > > > > > > > > > >
> > > > > > > > > > > > > Fixes: 306c54d0edb6 ("usb: hcd: Try MSI interrupts on
> > > > > > > > > > > > > PCI devices")
> > > > > >
> > > > > > > > > > Anyway, I applied this to pci/msi for v5.18 with the following
> > > > > > > > > > commit log:
> > > > > > > > > >
> > > > > > > > > >     PCI: Disable broken MSI on ATI SB600 USB adapters
> > > > > > > > > >
> > > > > > > > > >     Some ATI SB600 USB adapters advertise MSI, but MSI doesn't
> > > > > > > > > >     work if INTx is disabled.  Disable MSI on these adapters.
> > > > > > > > >
> > > > > > > > > But IIUC MSI is _not_ disabled. That's why I have issued this
> > > > > > > > > version of the patch with different commit message. Did I
> > > > > > > > > misunderstand something?
> > > > > > > >
> > > > > > > > Oh, right, of course.  Sorry, I was asleep at the wheel.
> > > > > > >
> > > > > > > Are you going to fix that?
> > > > > >
> > > > > > Yes, of course, I'll do something with the commit message after we
> > > > > > figure out how to handle PCI_COMMAND_INTX_DISABLE.
> > > > > >
> > > > > > > > I guess it's just that for these devices, we don't disable INTx
> > > > > > > > when enabling MSI.  I can't remember why we disable INTx when
> > > > > > > > enabling MSI, but it raises the question of whether it's better to
> > > > > > > > leave INTx enabled or to just disable use of MSI completely.
> > > > > > >
> > > > > > > It's required by specification to disable INTx if I read 6.1.4.3
> > > > > > > Enabling Operation correctly.
> > > > > >
> > > > > > Thanks for the reference; I was looking for something like that.  But
> > > > > > I don't think this section requires us to set
> > > > > > PCI_COMMAND_INTX_DISABLE.  For the benefit of folks without the spec,
> > > > > > PCIe r6.0, sec 6.1.4.3 says:
> > > > > >
> > > > > >   To maintain backward compatibility, the MSI Enable bit in the
> > > > > >   Message Control Register for MSI and the MSI-X Enable bit in the
> > > > > >   Message Control Register for MSI-X are each Clear by default (MSI
> > > > > >   and MSI-X are both disabled). System configuration software Sets one
> > > > > >   of these bits to enable either MSI or MSI-X, but never both
> > > > > >   simultaneously. Behavior is undefined if both MSI and MSI-X are
> > > > > >   enabled simultaneously. Software disabling either mechanism during
> > > > > >   active operation may result in the Function dropping pending
> > > > > >   interrupt conditions or failing to recognize new interrupt
> > > > > >   conditions. While enabled for MSI or MSI-X operation, a Function is
> > > > > >   prohibited from using INTx interrupts (if implemented) to request
> > > > > >   service (MSI, MSI-X, and INTx are mutually exclusive).
> > > > > >
> > > > > > The only *software* constraints I see are (1) software must never
> > > > > > enable both MSI and MSI-X simultaneously, and (2) if software disables
> > > > > > MSI or MSI-X during active operation, the Function may fail to
> > > > > > generate an interrupt when it should.
> > > > > >
> > > > > > I read the last sentence as a constraint on the *hardware*: if either
> > > > > > MSI or MSI-X is enabled, the Function is not allowed to use INTx,
> > > > > > regardless of the state of PCI_COMMAND_INTX_DISABLE.
> > > > > >
> > > > > > I searched the spec for "Interrupt Disable", looking for situations
> > > > > > where software might be *required* to set it, but I didn't see
> > > > > > anything.
> > > > > >
> > > > > > I suspect "Interrupt Disable" was intended to help the OS stop all
> > > > > > activity from a device during hot-plug or reconfiguration, as hinted
> > > > > > at in sec 6.4, "Device Synchronization":
> > > > > >
> > > > > >   The ability of the driver and/or system software to block new
> > > > > >   Requests from the device is supported by the Bus Master Enable,
> > > > > >   SERR# Enable, and Interrupt Disable bits in the Command register
> > > > > >   (Section 7.5.1.1.3) of each device Function, and other such control
> > > > > >   bits.
> > > > > >
> > > > > > So I'm trying to figure out why when enabling MSI we need to set
> > > > > > PCI_COMMAND_INTX_DISABLE for most devices, but it's safe to skip that
> > > > > > for these quirked devices.
> > > > >
> > > > > I guess it's wrong wording in the last paragraph. It's not safe, but it's
> > > > > _required_ since HW doesn't follow PCI specification that clearly says:
> > > > > "MSI, MSI-X, and INTx are mutually exclusive".
> > > >
> > > > I agree there's a defect in these SB600 devices.  My guess is that
> > > > PCI_COMMAND_INTX_DISABLE actually disables both INTx and MSI, when
> > > > it's only supposed to disable INTx.
> > > >
> > > > I'm pretty sure the spec doesn't actually require software to set
> > > > Interrupt Disable when enabling MSI, since MSI was added in PCI r2.2,
> > > > which included this text in sec 6.8.2:
> > > >
> > > >   System configuration software sets [the MSI Enable] bit to enable
> > > >   MSI. ...  Once enabled, a function is prohibited from using its
> > > >   INTx# pin (if implemented) to request service (MSI and INTx# are
> > > >   mutually exclusive).
> > > >
> > > > and Interrupt Disable was added later, in PCI r2.3, with no mention of
> > > > a connection with MSI.  All the specs from PCI r2.2 to PCIe r6.0
> > > > include the text above about not using INTx# if MSI or MSI-X is
> > > > enabled, but that's not the same as requiring software to set
> > > > Interrupt Disable.  Linux has set Interrupt Disable when enabling MSI
> > > > ever since MSI support was added [1], so I would hesitate to change
> > > > that even though I don't think it's required.
> > > 
> > > Thanks for diving into the history of the specification. What I learnt
> > > about any of the specifications is that it usually has a lot of
> > > implications that are only understandable (known) to the specification
> > > author(s). This gives a room of misinterpretation. In any case I
> > > usually apply my common sense denominator, so I try to go with the
> > > straight logic. In this case it seems to me that keeping both enabled
> > > is illogical and Linux does the right thing (means the author of the
> > > Linux kernel implementation is on the same page with me).
> > > 
> > > > What I don't like about PCI_DEV_FLAGS_MSI_INTX_DISABLE_BUG is that it
> > > > changes the generic code path in a sort of random way, i.e., this
> > > > device becomes yet another special case in how we handle Interrupt
> > > > Disable.
> > > >
> > > > What would you think about just setting pdev->no_msi instead, so we
> > > > don't try to use MSI at all on these devices?  I think that's what we
> > > > did before 306c54d0edb6.
> > > 
> > > Yes, we did. But why should we go this way if it already established a
> > > special case disregarding my patch(es)? If you want to do that you
> > > need to explain why other devices on the same chipset should enable
> > > MSI and what's wrong with enabling MSI on the USB devices. My
> > > understanding is that the MSI is a good thing to have due to
> > > performance benefits and taking into account other devices that have
> > > already been using it on the other devices of the same chipset tells
> > > me that's okay. Moreover, the reporter of the bug confirmed that MSI
> > > works for them after applying this quirk fix.
> > 
> > I agree that MSI is generally to be preferred over INTx.  In this
> > case, it's an old USB device and I don't think there's any real
> > performance benefit.
> 
> I tend to disagree. There are not only USB devices involved but other
> devices that may share the same INTx line(s). This is the case, for
> example, on EG20T. Dunno if ATI is one of them, but if so, it will
> support the idea of MSI.
> 
> > The problem with enabling MSI on these USB devices is that the generic
> > MSI code doesn't work correctly.
> 
> Can you elaborate this, please? Where is that code that doesn't work
> correctly?

On this device, pci_intx_for_msi() breaks MSI, so if we want to use
MSI, we have to set PCI_DEV_FLAGS_MSI_INTX_DISABLE_BUG to avoid the
breakage.

If a non-essential feature is broken, I think the cleanest path for
maintainability is to avoid using the feature.  In this case, MSI is
broken because Interrupt Disable seems to disable MSI as well as INTx.

If we enable MSI on these devices, we (which really means "I") have to
audit all the places that set Interrupt Disable to see if masking both
MSI and INTx will break anything.  This is the maintainability issue.

Most places that set PCI_COMMAND_INTX_DISABLE are in drivers for other
devices, which are not an issue, of course.  Some others:

  - pci_check_and_mask_intx(): used by vfio_intx_handler() (not a
    problem since this case uses INTx, not MSI) and the UIO
    irqhandler() (this is a potential problem for user-space drivers,
    but I don't know the implications).

  - vfio_pci_intx_mask() and similar VFIO paths: potentially an issue,
    but I guess guest drivers that use these are on their own and have
    to have their own knowledge of device quirks.

  - pci_dev_save_and_disable(): used in reset paths, so shouldn't be
    an issue.

  - xen_pcibk_get_interrupt_type(): decides interrupt type based on
    reading PCI_COMMAND_INTX_DISABLE and will return
    "INTERRUPT_TYPE_INTX | INTERRUPT_TYPE_MSI" when it should return
    "INTERRUPT_TYPE_MSI".  Will break msi_msix_flags_write(), which
    looks for an exact match, but I don't know the further
    implications.

  - Xen command_write(): No idea how to even figure out if this is an
    issue.

These all look like it *probably* wouldn't seriously break anything to
use PCI_DEV_FLAGS_MSI_INTX_DISABLE_BUG.  But the point is that it's a
lot of work for me to even make an educated guess, and "probably safe"
and "unknown benefit" don't make a strong case.

> > Would I prefer the same for all the other existing users of
> > PCI_DEV_FLAGS_MSI_INTX_DISABLE_BUG?  Yes.  Enough to change the status
> > quo, which would be a performance regression, and potentially break
> > something?  Probably not.
> 
> Why? What is the point of deliberate degrading of performance?

Sorry, I didn't make that clear enough.  I was trying to say that I
don't think it's a good idea to change *existing* users because it
would degrade performance and it might break something.

Bjorn

Powered by blists - more mailing lists