From mboxrd@z Thu Jan 1 00:00:00 1970 From: yangyingliang@huawei.com (Yang Yingliang) Date: Mon, 7 Sep 2015 10:54:52 +0800 Subject: [RFC PATCH v1 0/4] arm/arm64: fix a migrating irq bug when hotplug cpu In-Reply-To: <55EBF43E.8070501@linux.intel.com> References: <1441513421-8092-1-git-send-email-yangyingliang@huawei.com> <55EBF43E.8070501@linux.intel.com> Message-ID: <55ECFC7C.8060505@huawei.com> To: linux-arm-kernel@lists.infradead.org List-Id: linux-arm-kernel.lists.infradead.org On 2015/9/6 16:07, Jiang Liu wrote: > On 2015/9/6 12:23, Yang Yingliang wrote: >> Hi All, >> >> There is a bug: >> >> When cpu is disabled, all irqs will be migratged to another cpu. >> In some cases, a new affinity is different, it needed to be coppied >> to irq's affinity. But if the type of irq is LPI, it's affinity will >> not be coppied because of irq_set_affinity's return value. >> >> >> >> As Marc and Will suggested, I refactor the arm/arm64 migrating interrupts >> code and fix the migrating irq bug while cpu is offline. >> >> I'm trying let the core code do the migrating interrupts matter. kernel/irq/migration.c >> depends on CONFIG_GENERIC_PENDING_IRQ, so I make it selected by CONFIG_SMP and >> CONFIG_HOTPLUG_CPU and rename it to CONFIG_GENERIC_IRQ_MIGRATION for more general. >> When CONFIG_GENERIC_IRQ_MIGRATION is enabled, an interrupt whose state_use_accessors >> is not set with IRQD_MOVE_PCNTXT won't be migrated immediately in irq_set_affinity_locked(). >> So introduce irq_settings_set_move_pcntxt() helper to set the state in gic_irq_domain_map(). >> >> With the above preparation, move the migrating interrupts code into kernel/irq/migration.c >> and fix the bug by using irq_do_set_affinity(). > Hi Yingliang, > As we are going to move migrate_irqs() to generic kernel > code, and powerpc, metag, xtensa, sh, ia64 mn10300 also defines > migrate_irqs() too. It would be great if we could consolidate > all these. > And as we are going to refine these code, there's another > issue need attention. On x86, we need to allocate a CPU vector > if an irq is directed to a CPU. So there's possibility that > we run out of CPU vectors after CPU hot-removal. So we have a > mechanism to detect whether we will run out of CPU vector > after removing a CPU, and reject CPU hot-removal if that will > happen. > So the key point is, if we a need to allocate some sort > of resource on the target CPUs for an irq, we need two steps > when removing a CPU > 1) check whether resources are available after removing the CPU, > and reject CPU removal request if we ran out of resource > 2) fix irqs after hot-removing the CPU. > Thanks! > Gerry > On arm, as I know, it doesn't need extra resource for an irq. I am not sure other platform need this way besides x86. I think we could consolidate all migrate_irqs() later. I am not sure if it's good to do so big changing and modify other arch code in a patchset that supposed to fix a bug of arm.