Re: [PATCH 1/4] cpu: free cpu->tb_jmp_cache with RCU

[Richard Henderson]

On 1/8/23 08:39, Emilio Cota wrote:
> Fixes the appended use-after-free. The root cause is that
> during tb invalidation we use CPU_FOREACH, and therefore
> to safely free a vCPU we must wait for an RCU grace period
> to elapse.
>
> $ x86_64-linux-user/qemu-x86_64 tests/tcg/x86_64-linux-user/munmap-pthread
> =================================================================
> ==1800604==ERROR: AddressSanitizer: heap-use-after-free on address 
> 0x62d0005f7418 at pc 0x5593da6704eb bp 0x7f4961a7ac70 sp 0x7f4961a7ac60
> READ of size 8 at 0x62d0005f7418 thread T2
>      #0 0x5593da6704ea in tb_jmp_cache_inval_tb ../accel/tcg/tb-maint.c:244
>      #1 0x5593da6704ea in do_tb_phys_invalidate ../accel/tcg/tb-maint.c:290
>      #2 0x5593da670631 in tb_phys_invalidate__locked ../accel/tcg/tb-maint.c:306
>      #3 0x5593da670631 in tb_invalidate_phys_page_range__locked 
> ../accel/tcg/tb-maint.c:542
>      #4 0x5593da67106d in tb_invalidate_phys_range ../accel/tcg/tb-maint.c:614
>      #5 0x5593da6a64d4 in target_munmap ../linux-user/mmap.c:766
>      #6 0x5593da6dba05 in do_syscall1 ../linux-user/syscall.c:10105
>      #7 0x5593da6f564c in do_syscall ../linux-user/syscall.c:13329
>      #8 0x5593da49e80c in cpu_loop ../linux-user/x86_64/../i386/cpu_loop.c:233
>      #9 0x5593da6be28c in clone_func ../linux-user/syscall.c:6633
>      #10 0x7f496231cb42 in start_thread nptl/pthread_create.c:442
>      #11 0x7f49623ae9ff  (/lib/x86_64-linux-gnu/libc.so.6+0x1269ff)
>
> 0x62d0005f7418 is located 28696 bytes inside of 32768-byte region 
> [0x62d0005f0400,0x62d0005f8400)
> freed by thread T148 here:
>      #0 0x7f49627b6460 in __interceptor_free 
> ../../../../src/libsanitizer/asan/asan_malloc_linux.cpp:52
>      #1 0x5593da5ac057 in cpu_exec_unrealizefn ../cpu.c:180
>      #2 0x5593da81f851  (/home/cota/src/qemu/build/qemu-x86_64+0x484851)
>
> Signed-off-by: Emilio Cota <cota@braap.org>
> ---
>   accel/tcg/cpu-exec.c  | 1 -
>   cpu.c                 | 7 +++++++
>   include/hw/core/cpu.h | 3 +++
>   3 files changed, 10 insertions(+), 1 deletion(-)
>
> diff --git a/accel/tcg/cpu-exec.c b/accel/tcg/cpu-exec.c
> index 356fe348de..ca95d21528 100644
> --- a/accel/tcg/cpu-exec.c
> +++ b/accel/tcg/cpu-exec.c
> @@ -1070,7 +1070,6 @@ void tcg_exec_unrealizefn(CPUState *cpu)
>   #endif /* !CONFIG_USER_ONLY */
>   
>       tlb_destroy(cpu);
> -    g_free(cpu->tb_jmp_cache);

Can you simply use g_free_rcu here?

>   }
>   
>   #ifndef CONFIG_USER_ONLY
> diff --git a/cpu.c b/cpu.c
> index 4a7d865427..564200559f 100644
> --- a/cpu.c
> +++ b/cpu.c
> @@ -164,6 +164,12 @@ void cpu_exec_realizefn(CPUState *cpu, Error **errp)
>   #endif /* CONFIG_USER_ONLY */
>   }
>   
> +static void cpu_free_rcu(CPUState *cpu)
> +{
> +    /* .tb_jmp_cache is NULL except under TCG */
> +    g_free(cpu->tb_jmp_cache);
> +}
> +
>   void cpu_exec_unrealizefn(CPUState *cpu)
>   {
>   #ifndef CONFIG_USER_ONLY
> @@ -181,6 +187,7 @@ void cpu_exec_unrealizefn(CPUState *cpu)
>       }
>   
>       cpu_list_remove(cpu);
> +    call_rcu(cpu, cpu_free_rcu, rcu);

Certainly this seems wrong, exposing tb_jmp_cache beyond tcg.


r~