Re: [PATCH] PCI/MSI: Clear msi_desc::irq in the legacy teardown path

[sashiko-bot@kernel.org]

> pci_msi_teardown_msi_irqs() runs the legacy teardown and then calls
> msi_free_msi_descs(), which refuses to free a descriptor that is still
> associated with an interrupt:
> 
> 	/* Leak the descriptor when it is still referenced */
> 	if (WARN_ON_ONCE(msi_desc_match(desc, MSI_DESC_ASSOCIATED)))
> 		continue;
> 
> MSI_DESC_ASSOCIATED is msi_desc::irq being non-zero, and nothing in the
> legacy path clears it - neither the generic arch_teardown_msi_irqs() nor
> the arch_teardown_msi_irq() implementations on sparc and mips/octeon.
> Every teardown therefore leaks one descriptor per vector. The check is
> WARN_ON_ONCE, so only the first teardown after boot is visible and the
> warning understates how often this happens.
> 
> Before commit 9fb9eb4b59ac ("PCI/MSI: Let core code free MSI descriptors")
> free_msi_irqs() freed the descriptors unconditionally, so a stale
> msi_desc::irq was harmless. That commit moved the freeing into the core
> and added the precondition without satisfying it here.
> 
> powerpc overrides arch_teardown_msi_irqs() and clears msi_desc::irq
> itself, so it is unaffected; do the same in the generic implementation.
> 
> Reproduced on an UltraSPARC T7-1 (ixgbe, "ethtool -L <if> combined 4")
> and on an UltraSPARC T4-1 (igb, unbinding the PCI function). Verified
> fixed on the T7-1: repeated MSI-X teardown and setup is clean.
> 
> Fixes: 9fb9eb4b59ac ("PCI/MSI: Let core code free MSI descriptors")
> Closes: https://github.com/sparclinux/issues/issues/104
> Signed-off-by: Stian Halseth <stian@itx.no>

Sashiko has reviewed this patch and found no issues. It looks great!

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20260929111003.3707239-1-stian@itx.no?part=1