Re: Why set .suppress_bind_attrs even though .remove() implemented?

[Bjorn Helgaas]

[+to Johan for qcom]
[-cc Tom, email bounces]

On Thu, Jul 21, 2022 at 10:46:07PM +0200, Pali Rohár wrote:
> On Thursday 21 July 2022 14:54:33 Bjorn Helgaas wrote:
> > The j721e, kirin, tegra, and mediatek drivers all implement .remove().
> > 
> > They also set ".suppress_bind_attrs = true".  I think this means
> > bus_add_driver() will not create the "bind" and "unbind" sysfs
> > attributes for the driver that would allow users to users to manually
> > attach and detach devices from it.
> > 
> > Is there a reason for this, or should these drivers stop setting
> > .suppress_bind_attrs?
> 
> I have already asked this question during review of kirin driver:
> https://lore.kernel.org/linux-pci/20211031205527.ochhi72dfu4uidii@pali/
> 
> Microchip driver wanted to change its type from bool to tristate
> https://lore.kernel.org/linux-pci/20220420093449.38054-1-u.kleine-koenig@xxxxxxxxxxxxxx/t/#u
> and after discussion it seems that it is needed to do more work for this
> driver.
> 
> > For example, Pali and Ley Foon *did* stop setting .suppress_bind_attrs
> > when adding .remove() methods in these commits:
> > 
> >   0746ae1be121 ("PCI: mvebu: Add support for compiling driver as module")
> >   526a76991b7b ("PCI: aardvark: Implement driver 'remove' function and allow to build it as module")
> >   ec15c4d0d5d2 ("PCI: altera: Allow building as module")
> 
> I added it for both pci-mvebu.c and pci-aardvark.c. And just few days
> ago I realized why suppress_bind_attrs was set to true and remove method
> was not implemented.

With suppress_bind_attrs, the user can't manually unbind a device, so
we can't get to mvebu_pcie_remove() that way, but since mvebu is a
modular driver, I assume we can unload the module and *that* would
call mvebu_pcie_remove().  Right?

> Implementing remove method is not really simple, specially when pci
> controller driver implements also interrupt controller (e.g. for
> handling legacy interrupts).

Hmmm.  Based on your patches below, it looks like we need to call
irq_dispose_mapping() in some cases, but I'm very confused about
*which* cases.

I first thought it was for mappings created with irq_create_mapping(),
but pci-aardvark.c never calls that, so there must be more to it.

Currently only altera, iproc, mediatek-gen3, and mediatek call
irq_dispose_mapping() from their .remove() methods.  (They all call
irq_domain_remove() *before* irq_dispose_mapping().  Is that legal?
Your patches do irq_dispose_mapping() *first*.)

altera, mediatek-gen3, and mediatek call irq_dispose_mapping() on IRQs
that came from platform_get_irq().

qcom is a DWC driver, so all the IRQ stuff happens in
dw_pcie_host_init().  qcom_pcie_remove() does call
dw_pcie_host_deinit(), which calls irq_domain_remove(), but nobody
calls irq_dispose_mapping().

I'm thoroughly confused by all this.  But I suspect that maybe I
should drop the "make qcom modular" patch because it seems susceptible
to this problem:

  https://git.kernel.org/pub/scm/linux/kernel/git/helgaas/pci.git/commit/?h=pci/ctrl/qcom&id=41b68c2d097e

> Here are waiting fixup patches for pci-mvebu.c and pci-aardvark.c which
> fixes .remove callback. Without these patches calling 'rmmod driver' let
> dangling pointer in kernel which may cause random kernel crashes. See:
> 
> https://lore.kernel.org/linux-pci/20220709161858.15031-1-pali@xxxxxxxxxx/
> https://lore.kernel.org/linux-pci/20220711120626.11492-1-pali@xxxxxxxxxx/
> https://lore.kernel.org/linux-pci/20220711120626.11492-2-pali@xxxxxxxxxx/
> 
> So I would suggest to do more detailed review when adding .remove
> callback for pci controller driver (or when remove suppress_bind_attrs)
> and do more testings and checking if all IRQ mappings are disposed.

I'm not smart enough to do "more detailed review" because I don't know
what things to look for :)  Thanks for all your work in sorting out
these arcane details!

Bjorn