[PATCH 0/37] PCI/MSI: Enforce explicit IRQ vector management by removing devres auto-free

[Andy Shevchenko]

On Tue, Feb 24, 2026 at 12:09:37AM +0800, Shawn Lin wrote:
> 在 2026/02/23 星期一 23:50, Andy Shevchenko 写道:
> > On Mon, Feb 23, 2026 at 5:32 PM Shawn Lin <shawn.lin at rock-chips.com> wrote:
> > > 
> > > This patch series addresses a long-standing design issue in the PCI/MSI
> > > subsystem where the implicit, automatic management of IRQ vectors by
> > > the devres framework conflicts with explicit driver cleanup, creating
> > > ambiguity and potential resource management bugs.
> > > 
> > > ==== The Problem: Implicit vs. Explicit Management ====
> > > Historically, `pcim_enable_device()` not only manages standard PCI resources
> > > (BARs) via devres but also implicitly triggers automatic IRQ vector management
> > > by setting a flag that registers `pcim_msi_release()` as a cleanup action.
> > > 
> > > This creates an ambiguous ownership model. Many drivers follow a pattern of:
> > > 1. Calling `pci_alloc_irq_vectors()` to allocate interrupts.
> > > 2. Also calling `pci_free_irq_vectors()` in their error paths or remove routines.
> > > 
> > > When such a driver also uses `pcim_enable_device()`, the devres framework may
> > > attempt to free the IRQ vectors a second time upon device release, leading to
> > > a double-free. Analysis of the tree shows this hazardous pattern exists widely,
> > > while 35 other drivers correctly rely solely on the implicit cleanup.
> > 
> > Is this confirmed? What I read from the cover letter, this series was
> > only compile-tested, so how can you prove the problem exists in the
> > first place?
> 
> Yes, it's confirmed. My debug of a double free issue of a out-of-tree
> PCIe wifi driver which uses
> pcim_enable_device + pci_alloc_irq_vectors + pci_free_irq_vectors expose
> it. And we did have a TODO to cleanup this hybrid usage, targeted in
> this cycle[1] suggested by Philipp:

Okay, fair enough. I think this bit was missing in the cover letter.

> [1] https://git.kernel.org/pub/scm/linux/kernel/git/pci/pci.git/log/?h=msi

> > > ==== The Solution: Making Management Explicit ====
> > > This series enforces a clear, predictable model:
> > > 1.  New Managed API (Patch 1/37): Introduces pcim_alloc_irq_vectors() and
> > >      pcim_alloc_irq_vectors_affinity(). Drivers that desire devres-managed IRQ
> > >      vectors should use these functions, which set the is_msi_managed flag and
> > >      ensure automatic cleanup.
> > > 2.  Patches 2 through 36 convert each driver that uses pcim_enable_device() alongside
> > >      pci_alloc_irq_vectors() and relies on devres for IRQ vector cleanup to instead
> > >      make an explicit call to pcim_alloc_irq_vectors().
> > > 3.  Core Change (Patch 37/37): With the former cleanup, now modifies pcim_setup_msi_release()
> > >      to check only the is_msi_managed flag. This decouples automatic IRQ cleanup from
> > >      pcim_enable_device(). IRQ vectors allocated via pci_alloc_irq_vectors*()
> > >      are now solely the driver's responsibility to free with pci_free_irq_vectors().
> > > 
> > > With these changes, we clear ownership model: Explicit resource management eliminates
> > > ambiguity and follows the "principle of least surprise." New drivers choose one model and
> > > be consistent.
> > > - Use `pci_alloc_irq_vectors()` + `pci_free_irq_vectors()` for explicit control.
> > > - Use `pcim_alloc_irq_vectors()` for devres-managed, automatic cleanup.
> > 
> > Have you checked previous attempts? Why is your series better than those?
> 
> There seems not previous attempts.

Maybe we are looking to the different projects...

https://lore.kernel.org/all/?q=pcim_alloc_irq_vectors

> > > ==== Testing And Review ====
> > > 1. This series is only compiled test with allmodconfig.
> > > 2. Given the substantial size of this patch series, I have structured the mailing
> > >     to facilitate efficient review. The cover letter, the first patch and the last one will be sent
> > >     to all relevant mailing lists and key maintainers to ensure broad visibility and
> > >     initial feedback on the overall approach. The remaining subsystem-specific patches
> > >     will be sent only to the respective subsystem maintainers and their associated
> > >     mailing lists, reducing noise.

-- 
With Best Regards,
Andy Shevchenko