Re: BUG: KCSAN: data-race in do_page_fault / spectre_v4_enable_task_mitigation

[Will Deacon]

On Wed, Dec 21, 2022 at 03:54:36PM +0100, Anders Roxell wrote:
> Hey,
> 
> I'm building an allmodconfig kernel on yesterdays linux-next (tag:
> next-20221220) and I see a
> "BUG: KCSAN: data-race in do_page_fault / spectre_v4_enable_task_mitigation"
> when I boot up in QEMU. I ran the output via
> scripts/decode_stacktrace.sh and this is what I see:
> 
> 
> [ 2105.261121][  T154] ==================================================================
> [ 2105.266067][  T154] BUG: KCSAN: data-race in do_page_fault / spectre_v4_enable_task_mitigation
> [ 2105.271577][  T154] 
> [ 2105.273121][  T154] write to 0xffff8000210b3fb8 of 8 bytes by task 136 on cpu 0:
> [ 2105.277743][ T154] spectre_v4_enable_task_mitigation (/home/anders/src/kernel/next/arch/arm64/kernel/proton-pack.c:651 /home/anders/src/kernel/next/arch/arm64/kernel/proton-pack.c:664) 
> [ 2105.281802][ T154] __switch_to (/home/anders/src/kernel/next/arch/arm64/kernel/process.c:459 /home/anders/src/kernel/next/arch/arm64/kernel/process.c:532) 
> [ 2105.284670][ T154] __schedule (/home/anders/src/kernel/next/kernel/sched/core.c:5247 /home/anders/src/kernel/next/kernel/sched/core.c:6555) 
> [ 2105.287555][ T154] preempt_schedule_irq (/home/anders/src/kernel/next/arch/arm64/include/asm/irqflags.h:70 /home/anders/src/kernel/next/arch/arm64/include/asm/irqflags.h:98 /home/anders/src/kernel/next/kernel/sched/core.c:6868) 
> [ 2105.290857][ T154] arm64_preempt_schedule_irq (/home/anders/src/kernel/next/arch/arm64/kernel/entry-common.c:237) 
> [ 2105.294433][ T154] el1_interrupt (/home/anders/src/kernel/next/arch/arm64/kernel/entry-common.c:476 /home/anders/src/kernel/next/arch/arm64/kernel/entry-common.c:486) 
> [ 2105.297433][ T154] el1h_64_irq_handler (/home/anders/src/kernel/next/arch/arm64/kernel/entry-common.c:492) 
> [ 2105.300718][ T154] el1h_64_irq (/home/anders/src/kernel/next/arch/arm64/kernel/entry.S:580) 
> [ 2105.303497][ T154] arch_local_irq_restore (/home/anders/src/kernel/next/arch/arm64/include/asm/jump_label.h:21 /home/anders/src/kernel/next/arch/arm64/include/asm/irqflags.h:130) 
> [ 2105.306750][ T154] fs_reclaim_acquire (/home/anders/src/kernel/next/mm/page_alloc.c:4691) 
> [ 2105.310118][ T154] slab_pre_alloc_hook.constprop.0 (/home/anders/src/kernel/next/include/linux/sched/mm.h:272 /home/anders/src/kernel/next/mm/slab.h:720) 
> [ 2105.313966][ T154] slab_alloc_node.isra.0 (/home/anders/src/kernel/next/mm/slub.c:3434) 
> [ 2105.317343][ T154] __kmem_cache_alloc_lru (/home/anders/src/kernel/next/mm/slub.c:3469) 
> [ 2105.320659][ T154] kmem_cache_alloc (/home/anders/src/kernel/next/mm/slub.c:3477) 
> [ 2105.323673][ T154] getname_flags (/home/anders/src/kernel/next/fs/namei.c:139) 
> [ 2105.326703][ T154] getname (/home/anders/src/kernel/next/fs/namei.c:218) 
> [ 2105.329377][ T154] do_sys_openat2 (/home/anders/src/kernel/next/fs/open.c:1304) 
> [ 2105.332352][ T154] __arm64_sys_openat (/home/anders/src/kernel/next/fs/open.c:1326) 
> [ 2105.335573][ T154] el0_svc_common.constprop.0 (/home/anders/src/kernel/next/arch/arm64/kernel/syscall.c:38 /home/anders/src/kernel/next/arch/arm64/kernel/syscall.c:52 /home/anders/src/kernel/next/arch/arm64/kernel/syscall.c:142) 
> [ 2105.339272][ T154] do_el0_svc (/home/anders/src/kernel/next/arch/arm64/kernel/syscall.c:197) 
> [ 2105.342025][ T154] el0_svc (/home/anders/src/kernel/next/arch/arm64/kernel/entry-common.c:133 /home/anders/src/kernel/next/arch/arm64/kernel/entry-common.c:142 /home/anders/src/kernel/next/arch/arm64/kernel/entry-common.c:638) 
> [ 2105.344687][ T154] el0t_64_sync_handler (/home/anders/src/kernel/next/arch/arm64/kernel/entry-common.c:656) 
> [ 2105.348089][ T154] el0t_64_sync (/home/anders/src/kernel/next/arch/arm64/kernel/entry.S:584) 
> [ 2105.350998][  T154] 
> [ 2105.352567][  T154] read to 0xffff8000210b3fb8 of 8 bytes by task 154 on cpu 0:
> [ 2105.357117][ T154] do_page_fault (/home/anders/src/kernel/next/arch/arm64/mm/fault.c:517 /home/anders/src/kernel/next/arch/arm64/mm/fault.c:558) 
> [ 2105.360110][ T154] do_translation_fault (/home/anders/src/kernel/next/arch/arm64/mm/fault.c:695) 
> [ 2105.363400][ T154] do_mem_abort (/home/anders/src/kernel/next/arch/arm64/mm/fault.c:831) 
> [ 2105.366400][ T154] el0_ia (/home/anders/src/kernel/next/arch/arm64/kernel/entry-common.c:133 /home/anders/src/kernel/next/arch/arm64/kernel/entry-common.c:142 /home/anders/src/kernel/next/arch/arm64/kernel/entry-common.c:534) 
> [ 2105.369059][ T154] el0t_64_sync_handler (/home/anders/src/kernel/next/arch/arm64/kernel/entry-common.c:662) 
> [ 2105.372445][ T154] el0t_64_sync (/home/anders/src/kernel/next/arch/arm64/kernel/entry.S:584) 
> [ 2105.375404][  T154] 
> [ 2105.376935][  T154] no locks held by systemd/154.
> [ 2105.379935][  T154] irq event stamp: 385
> [ 2105.382448][ T154] hardirqs last enabled at (385): local_daif_restore (/home/anders/src/kernel/next/arch/arm64/include/asm/daifflags.h:71 (discriminator 1)) 
> [ 2105.388413][ T154] hardirqs last disabled at (384): el0t_64_sync_handler (/home/anders/src/kernel/next/arch/arm64/kernel/entry-common.c:662) 
> [ 2105.394436][ T154] softirqs last enabled at (352): fpsimd_restore_current_state (/home/anders/src/kernel/next/arch/arm64/kernel/fpsimd.c:264 /home/anders/src/kernel/next/arch/arm64/kernel/fpsimd.c:1780) 
> [ 2105.400932][ T154] softirqs last disabled at (350): fpsimd_restore_current_state (/home/anders/src/kernel/next/include/linux/bottom_half.h:20 /home/anders/src/kernel/next/arch/arm64/kernel/fpsimd.c:242 /home/anders/src/kernel/next/arch/arm64/kernel/fpsimd.c:1773) 
> [ 2105.407394][  T154] 
> [ 2105.408909][  T154] value changed: 0x0000000060000000 -> 0x0000000060001000
> [ 2105.413225][  T154] 
> [ 2105.414746][  T154] Reported by Kernel Concurrency Sanitizer on:
> [ 2105.426169][  T154] Hardware name: linux,dummy-virt (DT)
> [ 2105.429528][  T154] ==================================================================
> 
> 
> The prctl case, which only gets called on 'current'. However, assuming
> the kernel could be preempted while accessing current->pt_regs->pstate,
> and then it races against the task switch.
> 
> Any idea what happens and how to fix it?

I can't quite decipher this against mainline, as the line numbers above in
do_page_fault() seem to correlate with reads of the esr, which is either
a register or a stack read. I also think everything in the stacktraces
is running in the context of 'current', so I can't see how we could really
race here.

:/

Will