Re: kobject lifetime issues in blk-mq

From: Ming Lei
Date: Mon Nov 12 2018 - 04:44:22 EST

Next message: Chao Fan: "[PATCH v11 1/5] x86/boot: Add efi_get_rsdp_addr() to dig out RSDP from EFI table"
Previous message: Geert Uytterhoeven: "Re: [PATCH 1/3] staging: emxx_udc: Added static modifier to udc_controller"
In reply to: Ming Lei: "Re: kobject lifetime issues in blk-mq"
Next in thread: Guenter Roeck: "Re: kobject lifetime issues in blk-mq"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]

On Mon, Nov 12, 2018 at 05:20:51PM +0800, Ming Lei wrote:
> On Fri, Nov 09, 2018 at 12:35:18PM -0800, Guenter Roeck wrote:
> > Hi,
> >
> > Images with CONFIG_DEBUG_KOBJECT=y, CONFIG_DEBUG_KOBJECT_RELEASE=y
> > enabled report kobject lifetime issues when booting an image in qemu
> > using virtio-scsi-pci.
> >
> > Bisect points to two different commits, depending on the kernel
> > configuration.
> >
> > 1st configuration: defconfig plus:
> >
> > CONFIG_VIRTIO_BLK=y
> > CONFIG_SCSI_LOWLEVEL=y
> > CONFIG_SCSI_VIRTIO=y
> > CONFIG_VIRTIO_PCI=y
> > CONFIG_VIRTIO_MMIO=y
> > CONFIG_DEBUG_OBJECTS=y
> > CONFIG_DEBUG_OBJECTS_FREE=y
> > CONFIG_DEBUG_OBJECTS_TIMERS=y
> > CONFIG_DEBUG_KOBJECT=y
> > CONFIG_DEBUG_KOBJECT_RELEASE=y
> >
> > Bisecting this configuration points to commit b5b6e8c8d3b4 ("scsi:
> > virtio_scsi: fix IO hang caused by automatic irq vector affinity")
> > as the culprit. This commit enforces the use of blk-mq for virtio_scsi;
> > the bisect log is therefore a bit misleading.
> >
> > 2nd configuration: As above, plus:
> > CONFIG_SCSI_MQ_DEFAULT=y
> >
> > With this configuration, bisect points to commit 7ea5fe31c12d
> > ("blk-mq: make lifetime consitent between q/ctx and its kobject").
> >
> > The qemu command line used to reproduce the problem is as follows.
> > The qemu version does not seem to matter (I tested with qemu versions
> > 2.5, 2.11, and 3.0).
> >
> > qemu-system-x86_64 \
> > -kernel arch/x86/boot/bzImage \
> > -device virtio-scsi-pci,id=scsi \
> > -device scsi-hd,bus=scsi.0,drive=d0 \
> > -drive file=./wheezy.img,format=raw,if=none,id=d0 \
> > -snapshot \
> > -m 2G -smp 4 -enable-kvm \
> > -net user,host=10.0.2.10,hostfwd=tcp::10022-:22 -net nic \
> > -nographic -monitor none \
> > -no-reboot \
> > -append "console=ttyS0 root=/dev/sda earlyprintk=serial panic_on_warn=1 panic=-1"
> >
> > Root file system:
> > https://storage.googleapis.com/syzkaller/wheezy.img
> > or:
> > https://github.com/groeck/linux-build-test/blob/master/rootfs/x86_64/rootfs.ext2.gz
> >
> > It seems that any root file system can be used to reproduce the problem.
> > Also, the problem is not limited to virtio-scsi-pci. I also tried
> > am53c974, lsi53c810, and lsi53c895a (after enabling the respective
> > drivers), with the same result.
> >
> > Overall, this suggests that the problem is related to blk-mq and was
> > indeed introduced with commit 7ea5fe31c12d.
> >
> > For reference, I attached an actual crash log as well as a set of bisect
> > logs below.
> >
> > Unfortunately, my understanding of blk-mq is not good enough to find the
> > underlying problem and suggest a fix. Please let me know if there is
> > anything else I can do to help fixing the problem.
> >
> > Note that the problem was originally reported by syzbot running a test
> > on chromeos-4.14. It may well be that the problem has already been
> > reported and is being worked on. If so, please apologize the noise.
> >
> > Thanks,
> > Guenter
> >
> > ---
> > [ 8.700038] ------------[ cut here ]------------
> > [ 8.700376] ODEBUG: free active (active state 0) object type: timer_list hint: delayed_work_timer_fn+0x0/0x10
> > [ 8.701032] WARNING: CPU: 0 PID: 0 at lib/debugobjects.c:329 debug_print_object+0x6a/0x80
> > [ 8.701032] Kernel panic - not syncing: panic_on_warn set ...
> > [ 8.701032] CPU: 0 PID: 0 Comm: swapper/0 Not tainted 4.20.0-rc1 #30
> > [ 8.701032] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Ubuntu-1.8.2-1ubuntu1 04/01/2014
> > [ 8.701032] Call Trace:
> > [ 8.701032] <IRQ>
> > [ 8.701032] dump_stack+0x46/0x5b
> > [ 8.701032] panic+0xf3/0x246
> > [ 8.701032] ? debug_print_object+0x6a/0x80
> > [ 8.701032] __warn+0xeb/0xf0
> > [ 8.701032] ? debug_print_object+0x6a/0x80
> > [ 8.701032] ? debug_print_object+0x6a/0x80
> > [ 8.701032] report_bug+0xb1/0x120
> > [ 8.701032] fixup_bug.part.10+0x13/0x30
> > [ 8.701032] do_error_trap+0x8f/0xb0
> > [ 8.701032] do_invalid_op+0x31/0x40
> > [ 8.701032] ? debug_print_object+0x6a/0x80
> > [ 8.701032] invalid_op+0x14/0x20
> > [ 8.701032] RIP: 0010:debug_print_object+0x6a/0x80
> > [ 8.701032] Code: 8b 43 10 83 c2 01 8b 4b 14 89 15 c1 e2 78 01 4c 8b 45 00 4c
> > 89 e6 48 c7 c7 f0 ca 41 a6 48 8b 14 c5 a0 f7 24 a6 e8 06 8e cc ff <0f> 0b 5b 83
> > 05 70 47 1a 01 01 5d 41 5c c3 83 05 65 47 1a 01 01 c3
> > [ 8.701032] RSP: 0018:ffff89b9fda03e48 EFLAGS: 00010082
> > [ 8.701032] RAX: 0000000000000000 RBX: ffff89b9fd49d0c8 RCX: ffffffffa6646e18
> > [ 8.701032] RDX: 0000000000000001 RSI: 0000000000000082 RDI: ffffffffa6cc596c
> > [ 8.701032] RBP: ffffffffa664a5c0 R08: 79616c6564203a74 R09: 5f6b726f775f6465
> > [ 8.701032] R10: ffff89b9fda03f10 R11: 3178302f3078302b R12: ffffffffa6403e22
> > [ 8.701032] R13: 0000000000000001 R14: ffffffffa6d4d458 R15: ffff89b9fc4c8780
> > [ 8.701032] ? debug_print_object+0x6a/0x80
> > [ 8.701032] debug_check_no_obj_freed+0x1b8/0x1ea
> > [ 8.701032] ? rcu_process_callbacks+0x2a7/0x750
> > [ 8.701032] kmem_cache_free+0x6e/0x1a0
> > [ 8.701032] ? __blk_release_queue+0x150/0x150
> > [ 8.701032] rcu_process_callbacks+0x2a7/0x750
> > [ 8.701032] __do_softirq+0xf2/0x2c7
> > [ 8.701032] irq_exit+0xb7/0xc0
> > [ 8.701032] smp_apic_timer_interrupt+0x67/0x130
> > [ 8.701032] apic_timer_interrupt+0xf/0x20
> > [ 8.701032] </IRQ>
> > [ 8.701032] RIP: 0010:default_idle+0x12/0x140
> > [ 8.701032] Code: fe ff ff c6 43 08 00 fb eb b0 31 c0 eb b4 e8 65 c8 60 ff 90
> > 90 90 90 90 41 54 55 53 65 8b 2d 15 c4 3b 5a 66 66 66 66 90 fb f4 <65> 8b 2d 07
> > c4 3b 5a 66 66 66 66 90 5b 5d 41 5c c3 65 8b 05 f6 c3
> > [ 8.701032] RSP: 0018:ffffffffa6603e98 EFLAGS: 00000246 ORIG_RAX: ffffffffffffff13
> > [ 8.701032] RAX: ffffffffa5c52d10 RBX: 0000000000000000 RCX: 0000000000000000
> > [ 8.701032] RDX: 0000000000000001 RSI: 0000000000000000 RDI: 000000020605092a
> > [ 8.701032] RBP: 0000000000000000 R08: ffffffffcff7766a R09: ffff89b9fda20b00
> > [ 8.701032] R10: 0000000000000000 R11: 0000000000002400 R12: ffffffffa6611740
> > [ 8.701032] R13: 0000000000000000 R14: 0000000000000000 R15: ffffffffa6611740
> > [ 8.701032] ? __sched_text_end+0x5/0x5
> > [ 8.701032] do_idle+0x19c/0x230
> > [ 8.701032] cpu_startup_entry+0x14/0x20
> > [ 8.701032] start_kernel+0x491/0x4b1
> > [ 8.701032] secondary_startup_64+0xa4/0xb0
> > [ 8.701032] Kernel Offset: 0x24200000 from 0xffffffff81000000 (relocation range: 0xffffffff80000000-0xffffffffbfffffff)
> >
>
> That is simply caused by enabling DEBUG_KOBJECT_RELEASE itself, which
> tries to do kboject_release() after delaying a while.
>
> Wrt. this issue, q->mq_kobj is embedded in 'request_queue' whose
> instance is freed in release handler of q->kobj. So when one instance
> of 'request_queue' is freed, DEBUG_KOBJECT_RELEASE still expects that
> the q->mq_kobj is active.
>
> The release handler of q->mq_kobj is NOP actually, so in reality it won't
> be one issue at all.

Given we have removed legacy IO path completely, it should be fine to
remove q->mq_kobj and simply put everything under the kobj of queue,
especially any attributes under mq aren't marked as stable ABI.

Thanks,
Ming

Next message: Chao Fan: "[PATCH v11 1/5] x86/boot: Add efi_get_rsdp_addr() to dig out RSDP from EFI table"
Previous message: Geert Uytterhoeven: "Re: [PATCH 1/3] staging: emxx_udc: Added static modifier to udc_controller"
In reply to: Ming Lei: "Re: kobject lifetime issues in blk-mq"
Next in thread: Guenter Roeck: "Re: kobject lifetime issues in blk-mq"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]