| Message ID | 1414082970-20775-1-git-send-email-marc.zyngier@arm.com |
|---|---|
| State | Not Applicable |
| Delegated to: | David Miller |
| Headers | show |
On Thu, Oct 23, 2014 at 05:49:30PM +0100, Marc Zyngier wrote: > Commit 18dcf433f3de (AHCI: Optimize single IRQ interrupt processing) > switched the single IRQ case of the AHCI driver to use threaded > threaded interrupts. > > During this conversion, only the IRQF_SHARED flag was provided. The net > effect of this in the presence of level interrupts is that the > interrupt will not be masked during the execution of the treaded > handler, leading to a screaming interrupt if the thread is not > scheduled quickly enough, specially in error conditions: [...] > Not good. The culprit is a missing IRQF_ONESHOT flag, which addition > solves the problem entierly. The number of interrupts drops radically: I worried how many devices hanging off the same IRQ line in legacy mode would fail to initialize after this update due to IRQF_ONESHOT mismatch. Given that I did not observe nor anticipate this kind of problem and LKP test shows a perf drop I am thinking may be we need to revert 18dcf43 for now. Tejun? > [Booting Debian Jessie to a prompt] > Before fix: > 88: 3562 0 GIC 88 ahci-sunxi > After fix: > 88: 1992 0 GIC 88 ahci-sunxi > > Tested on a A20 board (ahci-sunxi). > > Cc: Alexander Gordeev <agordeev@redhat.com> > Cc: Tejun Heo <tj@kernel.org> > Cc: linux-ide@vger.kernel.org > Signed-off-by: Marc Zyngier <marc.zyngier@arm.com> > --- > drivers/ata/libahci.c | 2 +- > 1 file changed, 1 insertion(+), 1 deletion(-) > > diff --git a/drivers/ata/libahci.c b/drivers/ata/libahci.c > index 5eb61c9..49c649f 100644 > --- a/drivers/ata/libahci.c > +++ b/drivers/ata/libahci.c > @@ -2492,7 +2492,7 @@ static int ahci_host_activate_single_irq(struct ata_host *host, int irq, > return rc; > > rc = devm_request_threaded_irq(host->dev, irq, ahci_single_irq_intr, > - ahci_thread_fn, IRQF_SHARED, > + ahci_thread_fn, IRQF_SHARED | IRQF_ONESHOT, > dev_driver_string(host->dev), host); > if (rc) > return rc; > -- > 2.0.4 >
diff --git a/drivers/ata/libahci.c b/drivers/ata/libahci.c index 5eb61c9..49c649f 100644 --- a/drivers/ata/libahci.c +++ b/drivers/ata/libahci.c @@ -2492,7 +2492,7 @@ static int ahci_host_activate_single_irq(struct ata_host *host, int irq, return rc; rc = devm_request_threaded_irq(host->dev, irq, ahci_single_irq_intr, - ahci_thread_fn, IRQF_SHARED, + ahci_thread_fn, IRQF_SHARED | IRQF_ONESHOT, dev_driver_string(host->dev), host); if (rc) return rc;
Commit 18dcf433f3de (AHCI: Optimize single IRQ interrupt processing) switched the single IRQ case of the AHCI driver to use threaded threaded interrupts. During this conversion, only the IRQF_SHARED flag was provided. The net effect of this in the presence of level interrupts is that the interrupt will not be masked during the execution of the treaded handler, leading to a screaming interrupt if the thread is not scheduled quickly enough, specially in error conditions: [ 2.728051] ahci-sunxi 1c18000.sata: controller can't do PMP, turning off CAP_PMP [ 2.735578] ahci-sunxi 1c18000.sata: SSS flag set, parallel bus scan disabled [ 2.742792] ahci-sunxi 1c18000.sata: AHCI 0001.0100 32 slots 1 ports 3 Gbps 0x1 impl platform mode [ 2.751789] ahci-sunxi 1c18000.sata: flags: ncq sntf stag pm led clo only pio slum part ccc [ 2.761539] scsi host0: ahci_platform [ 2.765754] ata1: SATA max UDMA/133 mmio [mem 0x01c18000-0x01c18fff] port 0x100 irq 88 [...] [ 3.127977] ata1: SATA link down (SStatus 0 SControl 300) [...] [ 3.162035] Waiting for root device /dev/sda1... [ 3.163700] random: nonblocking pool is initialized [ 3.326593] irq 88: nobody cared (try booting with the "irqpoll" option) [ 3.333296] CPU: 0 PID: 0 Comm: swapper/0 Not tainted 3.18.0-rc1+ #3077 [ 3.339936] [<c002b318>] (unwind_backtrace) from [<c0026980>] (show_stack+0x20/0x24) [ 3.347681] [<c0026980>] (show_stack) from [<c067dd18>] (dump_stack+0x9c/0xd4) [ 3.354904] [<c067dd18>] (dump_stack) from [<c007ac08>] (__report_bad_irq+0x38/0xd4) [ 3.362643] [<c007ac08>] (__report_bad_irq) from [<c007b1dc>] (note_interrupt+0x280/0x2d0) [ 3.370903] [<c007b1dc>] (note_interrupt) from [<c0078a70>] (handle_irq_event_percpu+0x210/0x27c) [ 3.379771] [<c0078a70>] (handle_irq_event_percpu) from [<c0078b28>] (handle_irq_event+0x4c/0x6c) [ 3.388636] [<c0078b28>] (handle_irq_event) from [<c007b860>] (handle_fasteoi_irq+0xbc/0x1a0) [ 3.397156] [<c007b860>] (handle_fasteoi_irq) from [<c0077f5c>] (generic_handle_irq+0x3c/0x4c) [ 3.405764] [<c0077f5c>] (generic_handle_irq) from [<c0078220>] (__handle_domain_irq+0x6c/0xb4) [ 3.414456] [<c0078220>] (__handle_domain_irq) from [<c0008754>] (gic_handle_irq+0x34/0x6c) [ 3.422802] [<c0008754>] (gic_handle_irq) from [<c0027540>] (__irq_svc+0x40/0x74) [ 3.430274] Exception stack(0xc09c1df8 to 0xc09c1e40) [ 3.435321] 1de0: 00000001 c0a27ec0 [ 3.443491] 1e00: 00000000 00000000 c09c0028 0000001e 00000282 00000000 ee00c400 c0a239f8 [ 3.451660] 1e20: c09c0000 c09c1ea4 c0a27ec0 c09c1e40 00200000 c003c7cc 20000113 ffffffff [ 3.459835] [<c0027540>] (__irq_svc) from [<c003c7cc>] (__do_softirq+0xbc/0x34c) [ 3.467229] [<c003c7cc>] (__do_softirq) from [<c003cd44>] (irq_exit+0xbc/0x104) [ 3.474536] [<c003cd44>] (irq_exit) from [<c0078224>] (__handle_domain_irq+0x70/0xb4) [ 3.482362] [<c0078224>] (__handle_domain_irq) from [<c0008754>] (gic_handle_x74) [ 3.498169] Exception stack(0xc09c1f10 to 0xc09c1f58) [ 3.503216] 1f00: ee7ca310 00000000 00000ab4 c0035560 [ 3.511386] 1f20: c09c0000 00000000 c0a23f84 c09c8474 c09c84bc c09c0000 c0688bf8 c09c1f64 [ 3.519554] 1f40: c09c1f68 c09c1f58 c00235d8 c00235dc 60000113 ffffffff [ 3.526165] [<c0027540>] (__irq_svc) from [<c00235dc>] (arch_cpu_idle+0x48/0x4c) [ 3.533557] [<c00235dc>] (arch_cpu_idle) from [<c00701fc>] (cpu_startup_entry+0x13c/0x27c) [ 3.541816] [<c00701fc>] (cpu_startup_entry) from [<c067af1c>] (rest_init+0x94/0x98) [ 3.549557] [<c067af1c>] (rest_init) from [<c093dd38>] (start_kernel+0x3f0/0x3fc) [ 3.557036] [<c093dd38>] (start_kernel) from [<40008084>] (0x40008084) [ 3.563555] handlers: [ 3.565830] [<c03a1a80>] ahci_single_irq_intr threaded [<c03a223c>] ahci_thread_fn [ 3.573415] Disabling IRQ #88 [ 3.576532] ata1: exception Emask 0x10 SAct 0x0 SErr 0x4050002 action 0xe frozen [ 3.583977] ata1: irq_stat 0x00400040, connection status changed [ 3.590006] ata1: SError: { RecovComm PHYRdyChg CommWake DevExch } [ 3.596209] ata1: hard resetting link [...] Not good. The culprit is a missing IRQF_ONESHOT flag, which addition solves the problem entierly. The number of interrupts drops radically: [Booting Debian Jessie to a prompt] Before fix: 88: 3562 0 GIC 88 ahci-sunxi After fix: 88: 1992 0 GIC 88 ahci-sunxi Tested on a A20 board (ahci-sunxi). Cc: Alexander Gordeev <agordeev@redhat.com> Cc: Tejun Heo <tj@kernel.org> Cc: linux-ide@vger.kernel.org Signed-off-by: Marc Zyngier <marc.zyngier@arm.com> --- drivers/ata/libahci.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)