[PATCH RFC 00/77] Re-design MSI/MSI-X interrupts enablement pattern

[Ben Hutchings]

On Fri, 2013-10-04 at 10:29 +0200, Alexander Gordeev wrote:
> On Thu, Oct 03, 2013 at 11:49:45PM +0100, Ben Hutchings wrote:
> > On Wed, 2013-10-02 at 12:48 +0200, Alexander Gordeev wrote:
> > > This update converts pci_enable_msix() and pci_enable_msi_block()
> > > interfaces to canonical kernel functions and makes them return a
> > > error code in case of failure or 0 in case of success.
> > [...]
> > 
> > I think this is fundamentally flawed: pci_msix_table_size() and
> > pci_get_msi_cap() can only report the limits of the *device* (which the
> > driver usually already knows), whereas MSI allocation can also be
> > constrained due to *global* limits on the number of distinct IRQs.
> 
> Even the current implementation by no means addresses it. Although it
> might seem a case for architectures to report the number of IRQs available
> for a driver to retry, in fact they all just fail. The same applies to
> *any* other type of resource involved: irq_desc's, CPU interrupt vector
> space, msi_desc's etc. No platform cares about it and just bails out once
> a constrain met (please correct me if I am wrong here). Given that Linux
> has been doing well even on embedded I think we should not change it.
>
> The only exception to the above is pSeries platform which takes advantage
> of the current design (to implement MSI quota). There are indications we
> can satisfy pSeries requirements, but the design proposed in this RFC
> is not going to change drastically anyway. The start of the discusstion
> is here: https://lkml.org/lkml/2013/9/5/293

All I can see there is that Tejun didn't think that the global limits
and positive return values were implemented by any architecture.  But
you have a counter-example, so I'm not sure what your point is.

It has been quite a while since I saw this happen on x86.  But I just
checked on a test system running RHEL 5 i386 (Linux 2.6.18).  If I ask
for 16 MSI-X vectors on a device that supports 1024, the return value is
8, and indeed I can then successfully allocate 8.

Now that's going quite a way back, and it may be that global limits
aren't a significant problem any more.  With the x86_64 build of RHEL 5
on an identical system, I can allocate 16 or even 32, so this is
apparently not a hardware limit in this case.

> > Currently pci_enable_msix() will report a positive value if it fails due
> > to the global limit.  Your patch 7 removes that.  pci_enable_msi_block()
> > unfortunately doesn't appear to do this.
> 
> pci_enable_msi_block() can do more than one MSI only on x86 (with IOMMU),
> but it does not bother to return positive numbers, indeed.
> 
> > It seems to me that a more useful interface would take a minimum and
> > maximum number of vectors from the driver.  This wouldn't allow the
> > driver to specify that it could only accept, say, any even number within
> > a certain range, but you could still leave the current functions
> > available for any driver that needs that.
> 
> Mmmm.. I am not sure I am getting it. Could you please rephrase?

Most drivers seem to either:
(a) require exactly a certain number of MSI vectors, or
(b) require a minimum number of MSI vectors, usually want to allocate
more, and work with any number in between

We can support drivers in both classes by adding new allocation
functions that allow specifying a minimum (required) and maximum
(wanted) number of MSI vectors.  Those in class (a) would just specify
the same value for both.  These new functions can take account of any
global limit or allocation policy without any further changes to the
drivers that use them.

The few drivers with more specific requirements would still need to
implement the currently recommended loop, using the old allocation
functions.

Ben.

-- 
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.