Re: [PATCH 01/17] PCI: Add concurrency safe clear_and_set variants for LNKCTL{,2}

["Ilpo Järvinen" <ilpo.jarvinen@linux.intel.com>]

[-- Attachment #1: Type: text/plain, Size: 4118 bytes --]

On Sun, 14 May 2023, Lukas Wunner wrote:

> On Fri, May 12, 2023 at 11:25:32AM +0300, Ilpo Järvinen wrote:
> > On Thu, 11 May 2023, Lukas Wunner wrote:
> > > On Thu, May 11, 2023 at 10:55:06AM -0500, Bjorn Helgaas wrote:
> > > > I didn't see the prior discussion with Lukas, so maybe this was
> > > > answered there, but is there any reason not to add locking to
> > > > pcie_capability_clear_and_set_word() and friends directly?
> > > > 
> > > > It would be nice to avoid having to decide whether to use the locked
> > > > or unlocked versions.
> > > 
> > > I think we definitely want to also offer lockless accessors which
> > > can be used in hotpaths such as interrupt handlers if the accessed
> > > registers don't need any locking (e.g. because there are no concurrent
> > > accesses).
> > > 
> > > So the relatively lean approach chosen here which limits locking to
> > > Link Control and Link Control 2, but allows future expansion to other
> > > registers as well, seemed reasonable to me.
> > 
> > I went through every single use of these functions in the mainline tree 
> > excluding LNKCTL/LNKCTL2 ones which will be having the lock anyway:
> > 
> > - pcie_capability_clear_and_set_*
> > - pcie_capability_set_*
> > - pcie_capability_clear_*
> 
> We're also performing RMW through pcie_capability_read_word() +
> pcie_capability_write_word() combos, see drivers/pci/hotplug/pciehp_hpc.c
> for examples.

That's why I said there could be other RMW operations outside of what
I carefully looked at. It, however, does not mean I didn't take any look 
at those.

But since brought it up, lets go through this case with
drivers/pci/hotplug/pciehp_hpc.c, it won't change anything:

All PCI_EXP_SLTSTA ones looked not real RMW but ACK bits type of writes
(real RMW = preverse other bits vs ACK write = other bits are written as 
zeros). Using RMW accessors would need an odd construct such as this
(and pcie_capability_set/clear_word() would be plain wrong):
	pcie_capability_clear_and_set_word(dev, PCI_EXP_SLTSTA,
					   ~PCI_EXP_SLTSTA_CC,
					   PCI_EXP_SLTSTA_CC);

PCI_EXP_SLTCTL write is protected by a mutex, it doesn't look something 
that matches your initial concern about "hot paths (e.g. interrupt 
handlers)".

In general, outside of drivers/pci/hotplug there are not that many 
capability writes (beyond LNKCTL/LNKCTL2 and now also RTCTL). None of 
those seem hot paths.

> > Do you still feel there's a need to differentiate this per capability 
> > given all the information above?
> 
> What I think is unnecessary and counterproductive is to add wholesale
> locking of any access to the PCI Express Capability Structure.
> 
> It's fine to have a single spinlock, but I'd suggest only using it
> for registers which are actually accessed concurrently by multiple
> places in the kernel.

While it does feel entirely unnecessary layer of complexity to me, it would 
be possible to rename the original pcie_capability_clear_and_set_word() to 
pcie_capability_clear_and_set_word_unlocked() and add this into 
include/linux/pci.h:

static inline int pcie_capability_clear_and_set_word(struct pci_dev *dev,
					int pos, u16 clear, u16 set)
{
	if (pos == PCI_EXP_LNKCTL || pos == PCI_EXP_LNKCTL2 ||
	    pos == PCI_EXP_RTCTL)
		pcie_capability_clear_and_set_word_locked(...);
	else
		pcie_capability_clear_and_set_word_unlocked(...);
}

It would keep the interface exactly the same but protect only a selectable 
set of registers. As pos is always a constant, the compiler should be able 
to optimize all the dead code away.

Would that be ok then?

-- 
 i.


> > spinlock + irq / work drivers/pci/pcie/pme.c: pcie_capability_set_word(dev, PCI_EXP_RTCTL,
> > spinlock + irq / work drivers/pci/pcie/pme.c: pcie_capability_clear_word(dev, PCI_EXP_RTCTL,
> [...]
> > What's more important though, isn't it possible that AER and PME RMW
> > PCI_EXP_RTCTL at the same time so it would need this RMW locking too 
> > despite the pme internal spinlock?
> 
> Yes that looks broken, so RTCTL would be another register besides
> LNKCTL and LNKCTL2 that needs protection, good catch.
> 
> Thanks,
> 
> Lukas
>